Abnormality sign detection system and abnormality sign detection method

Abstract

To provide an abnormality sign detection system that is capable of detecting abnormal signs in a target device with high accuracy.SOLUTION: An abnormality sign detection system 100 includes a processing unit 10 that detects abnormality signs of a target facility based on a classification model using a data set including processed sensor data in which detection values of a plurality of sensors of the target facility are subjected to predetermined pre-processing and abnormality determination data indicating the presence or absence of an abnormality sign of the target facility as training data. For any date during a given past period in which an inspector's inspection of the target facility was normal, the processing unit 10 regards the target facility as being normal at each time on that date, and generates abnormality determination data.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to an abnormality sign detection system and the like. [Background technology]

[0002] As a technology for detecting signs of abnormality in water treatment equipment and the like, for example, the technology described in Patent Document 1 is known. That is, Patent Document 1 describes that "a dispersion state is calculated from a group of differential pressure measurement values ​​in a second period, and whether or not the submerged membrane separation device is abnormal is determined based on the dispersion state." [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2022-100265 Summary of the Invention [Problem to be solved by the invention]

[0004] In the technology described in Patent Document 1, the presence or absence of an abnormality in a submerged membrane separation device is determined based on the dispersion state calculated from a group of differential pressure measurement values, but there is room for further improvement in the accuracy of detecting signs of abnormality.

[0005] Therefore, an object of the present disclosure is to provide an anomaly sign detection system etc. capable of detecting anomaly signs in target equipment with high accuracy. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, the anomaly sign detection system according to the present disclosure includes a processing unit that detects signs of anomaly in the target equipment based on a classification model using as training data a dataset including processed sensor data in which a predetermined preprocessing has been applied to the detection values ​​of multiple sensors of the target equipment, and anomaly judgment data indicating the presence or absence of signs of anomaly in the target equipment, and the processing unit is configured to generate the anomaly judgment data by assuming that the target equipment was normal at each time on any date during a predetermined past period on which an inspector's inspection of the target equipment was normal. Effect of the Invention

[0007] According to the present disclosure, it is possible to provide an abnormality sign detection system and the like that is capable of detecting abnormality signs in target equipment with high accuracy. [Brief description of the drawings]

[0008] [Figure 1] 1 is an explanatory diagram of a water treatment facility that is an example of a target facility of an abnormality sign detection system according to an embodiment. FIG. [Diagram 2] 1 is a functional block diagram of an abnormality sign detection system according to an embodiment. [Diagram 3] 2 is an explanatory diagram showing an example of a sensor database of the anomaly sign detection system according to the embodiment; FIG. [Figure 4] 2 is an explanatory diagram showing an example of an inspection record database of the anomaly sign detection system according to the embodiment; FIG. [Diagram 5] 4 is a flowchart of a process related to learning of the anomaly sign detection system according to the embodiment. [Figure 6] 4 is a flowchart relating to pre-processing of the abnormality sign detection system according to the embodiment. [Figure 7] 1 is an explanatory diagram showing an example of a processed sensor database of an anomaly sign detection system according to an embodiment; [Figure 8] 1 is an explanatory diagram showing an example of a processed inspection record database of an anomaly sign detection system according to an embodiment; FIG. [Figure 9]4 is a flowchart illustrating an abnormality detection process of the abnormality sign detection system according to the embodiment. [Figure 10A] 1 is an explanatory diagram regarding the processing performed when the inspection record results in an abnormality in the abnormality sign detection system according to the embodiment. FIG. [Figure 10B] 13 is an explanatory diagram regarding the processing performed when the inspection record results are normal in the anomaly sign detection system according to the embodiment. FIG. [Figure 11] 2 is an explanatory diagram showing an example of an anomaly determination database of the anomaly sign detection system according to the embodiment; FIG. [Figure 12] 11 is a flowchart relating to selection of input data for the anomaly sign detection system according to the embodiment. [Figure 13] 11 is a flowchart relating to classification model processing of the anomaly sign detection system according to the embodiment. [Figure 14] 4 is a flowchart of a process relating to inference of the abnormality sign detection system according to the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] <Embodiment> In the following, a water treatment facility 30 (see FIG. 1) that is an example of the "target facility" will be briefly described, and then the abnormality sign detection system 100 (see FIG. 2) will be described in detail.

[0010] <Water treatment facility configuration> FIG. 1 is an explanatory diagram of a water treatment facility 30 which is an example of a target facility of the abnormality sign detection system according to the embodiment. The water treatment facility 30 shown in Fig. 1 is a facility that performs a predetermined treatment on raw water (e.g., water containing organic matter) to extract organic treated water (water from which organic matter has been treated). As shown in Fig. 1, the water treatment facility 30 includes a pH adjustment tank 31, an aeration tank 32, a settling tank 33, an organic treated water tank 34, a sludge storage tank 35, an agitation tank 36, a dehydrator 37, and a conveyor 38.

[0011] In the pH adjustment tank 31, the water to be treated, which is obtained by adding a predetermined chemical to the raw water, is stirred by the stirrer 31a, whereby the pH of the water to be treated is adjusted to a predetermined pH, such as neutral or weakly acidic.

[0012] The aeration tank 32 has a function of purifying the water to be treated using activated sludge containing aerobic microorganisms, and is connected to the downstream side of the pH adjustment tank 31 via a pipe. As shown in FIG. 1, a blower 32a and an aeration pipe 32b are installed in the aeration tank 32. The blower 32a is a device that supplies oxygen to the aeration pipe 32b. The aeration pipe 32b is a pipe for releasing the oxygen supplied from the blower 32a into the tank as fine bubbles, and is immersed in the water to be treated. The supply of oxygen to the water to be treated via the aeration pipe 32b promotes the decomposition of organic matter by the aerobic microorganisms. As a result, flocs are formed as clumps of aerobic microorganisms and the like.

[0013] The settling tank 33 has a function of separating flocs from the water to be treated in which organic matter has been decomposed, and is connected via piping to the downstream side of the aeration tank 32. A rake 33a for collecting the flocs is installed at the bottom of the settling tank 33. The flocs that have settled in the settling tank 33 are collected by the rake 33a as the motor 33b is driven, and some of them are returned to the aeration tank 32 via the piping.

[0014] The organic treatment water tank 34 stores the supernatant liquid (organic treatment water from which organic matter has been treated) after treatment in the settling tank 33, and is connected via a pipe to the upper part of the settling tank 33. The organic treatment water stored in the organic treatment water tank 34 is discharged in a predetermined manner.

[0015] The sludge storage tank 35 stores the flocks separated in the settling tank 33, and is connected to the lower part of the settling tank 33 via piping. The agitation tank 36 has a function of agitating the flocks flowing in from the sludge storage tank 35 with an agitator 36a, and is connected to the lower part of the sludge storage tank 35 via piping. The dehydrator 37 is a device that dehydrates the flocks agitated in the agitation tank 36. The dehydrated flocks are transported by a transporter 38 and then subjected to a predetermined treatment. Although not shown, predetermined sensors are installed in each tank and piping of the water treatment facility 30.

[0016] Moreover, the configuration of the water treatment facility 30 shown in Fig. 1 is an example, and is not limited thereto. For example, a membrane tank (not shown) having a membrane module (not shown) may be provided instead of the settling tank 33 shown in Fig. 1. The membrane module is a filtration membrane for separating sludge from permeate. For example, a microfiltration membrane (MF membrane) or an ultrafiltration membrane (UF membrane) is used as such a membrane module.

[0017] Furthermore, the "target facility" for abnormality sign detection is not limited to the water treatment facility 30. For example, the "target facility" for abnormality sign detection may be a manufacturing plant, a chemical plant, a power generation facility, a substation facility, a power distribution facility, a railway facility, a communication facility, an air conditioning facility, a medical facility, a gas facility, a water facility, or a computer facility.

[0018] 1 is a device that acquires predetermined sensor data, and is connected to each of a plurality of sensors (not shown) of the water treatment facility 30. The computer 50 executes predetermined processing (learning, inference, etc.) based on the sensor data acquired via the PLC 40.

[0019] FIG. 2 is a functional block diagram of the abnormality sign detection system 100. The abnormality sign detection system 100 shown in Fig. 2 is a system for detecting abnormality signs in a target facility (for example, water treatment facility 30: see Fig. 1). Note that an "abnormality sign" is a precursor to the occurrence of an abnormality in the target facility. The abnormality sign detection system 100 may be configured with one computer 50 (see Fig. 1), or may be configured with multiple computers (servers, etc.) that are capable of communicating with each other.

[0020] The abnormality sign detection system 100 includes a processing unit 10 shown in Fig. 2 as well as a storage unit (not shown) in which predetermined programs and databases are stored. This storage unit includes a non-volatile memory such as a read only memory (ROM) and a volatile memory such as a random access memory (RAM) and a register.

[0021] 2 is, for example, a CPU (Central Processing Unit), which reads and develops a program stored in a storage unit (not shown) to execute a predetermined process. Specifically, the processing unit 10 detects a sign of abnormality in the water treatment facility 30 based on a classification model using as training data a data set including processed sensor data in which detection values ​​of a plurality of sensors (not shown) in the water treatment facility 30 (target facility: see FIG. 1) have been subjected to a predetermined pre-processing, and abnormality determination data indicating the presence or absence of a sign of abnormality in the water treatment facility 30.

[0022] As shown in Fig. 2, the processing unit 10 includes an input data acquisition unit 11, a preprocessing unit 12, an abnormality determination unit 13, an input data selection unit 14, and a classification model unit 15. The databases stored in the storage unit (not shown) include a sensor database 21, an inspection record database 22, a processed sensor database 23, and a processed inspection record database 24, as well as an abnormality determination database 25, a classification model database 26, and a detection result database 27. Of these multiple databases, the sensor database 21 (see Fig. 3) and the inspection record database 22 (see Fig. 4) will first be described.

[0023] FIG. 3 is an explanatory diagram showing an example of the sensor database 21. As shown in FIG. 3 stores detection values ​​of a plurality of sensors (not shown) of the water treatment facility 30 (see FIG. 1). That is, the detection values ​​of the plurality of sensors are acquired by the PLC 40 (see FIG. 1) and stored in the sensor database 21 in association with the date and time of acquisition and the identification information of the sensor.

[0024] In the example of Fig. 3, the detection values ​​of the suction pressure and transmembrane pressure every 10 minutes, as well as the treatment flow rate (flow rate of treated water per unit time) of the water treatment facility 30, are stored in the sensor database 21 in association with the date and time (date and time). The suction pressure is the pressure when the water to be treated is sucked toward the membrane module (not shown). The transmembrane pressure is the pressure difference between the inside and outside of the membrane module (not shown). For example, as the membrane module becomes more clogged, the transmembrane pressure tends to increase.

[0025] For convenience, three types of sensor data, namely, processing flow rate, suction pressure, and transmembrane pressure, are shown in FIG. 3, but sensor data from tens, hundreds, or thousands of sensors (not shown) is usually acquired from time to time. The state quantities detected are not limited to those shown in FIG. 3. For example, state quantities such as a predetermined mass, time, temperature, concentration, amount of heat, pH value, current, and voltage may be detected. The period for detecting the predetermined state quantity is not limited to 10 minutes, and may be, for example, 0.1 seconds, 1 second, or 1 minute. The sampling period for detecting the state quantity may differ depending on the sensor.

[0026] FIG. 4 is an explanatory diagram showing an example of the inspection record database 22. As shown in FIG. In the inspection record database 22 shown in FIG. 4, the results of inspection records of the water treatment facility 30 (see FIG. 1) by an inspector are stored as inspection record data. The example in FIG. 4 shows a case where the water treatment facility 30 is inspected once a day. For example, the result of the inspection record of the water treatment facility 30 on January 1st is "abnormal", and the result of the inspection record on January 2nd is "normal". Although not shown, when the result of the inspection record is "abnormal", data such as the location of the abnormality, the type of abnormality, and the cause of the abnormality are input by the inspector and associated with the inspection record data. It is assumed that such inspection of the water treatment facility 30 (see FIG. 1) is mainly performed by an experienced inspector.

[0027] The input data acquiring unit 11 shown in Fig. 2 acquires predetermined input data. That is, the input data acquiring unit 11 acquires sensor data for a predetermined period from the sensor database 21 and also acquires inspection record data for a predetermined period from the inspection record database 22 as input data.

[0028] The preprocessing unit 12 performs a predetermined preprocessing on the sensor data and the inspection record data acquired by the input data acquiring unit 11. Here, the "preprocessing" refers to a preliminary process for making it possible to perform a regression analysis or the like with the inspection record data as a response variable and the sensor data as an explanatory variable. The processed sensor database 23 stores the processed sensor data after the preprocessing has been performed. The processed inspection record database 24 stores the processed inspection record data after the preprocessing has been performed.

[0029] The anomaly determination unit 13 calculates the degree of anomaly of the processed sensor data based on the Hotelling theory, and determines the presence or absence of anomaly signs at each date and time based on the result of comparing the degree of anomaly with a predetermined threshold. The Hotelling theory will be described later. The abnormality determination database 25 stores the results of processing by the abnormality determination unit 13. That is, the abnormality determination database 25 stores an abnormality determination data set (data set) in which the processed sensor data, the degree of abnormality, and the presence or absence of an abnormality sign in the water treatment facility 30 (see FIG. 1) are associated with each other.

[0030] The input data selection unit 14 identifies a combination of multiple types of processed sensor data (explanatory variables) that has a relatively high degree of association with the processed inspection record data (objective variable) in the anomaly determination data set based on a stepwise method. The stepwise method is a method for identifying a combination of multiple types of explanatory variables that has a high degree of association with a predetermined objective variable.

[0031] The classification model unit 15 learns a predetermined classification model based on a random forest technique using the anomaly determination data set (data set) identified by the input data selection unit 14 as training data. Note that the random forest will be described later.

[0032] The classification model database 26 stores predetermined classification models that are learning results of the classification model unit 15. The detection result database 27 stores the detection results of abnormality signs based on the classification models.

[0033] <Processing of abnormality prediction detection system> The processing of the anomaly sign detection system 100 can be broadly divided into "learning" and "inference." Here, "learning" refers to generating a classification model for determining the presence or absence of an anomaly sign in the water treatment facility 30 by machine learning. Also, "inference" refers to determining the presence or absence of an anomaly sign in the water treatment facility 30 based on predetermined sensor data and a classification model. Below, we will first explain the "learning" processing of the anomaly sign detection system (see FIG. 5), and then explain the "inference" processing (see FIG. 14).

[0034] FIG. 5 is a flowchart of a process related to learning of the anomaly sign detection system (also see FIG. 2 as appropriate). In step S101, the processing unit 10 acquires input data (i.e., sensor data and inspection record data) by the input data acquisition unit 11. For example, the processing unit 10 acquires sensor data for the most recent year from the sensor database 21, and acquires inspection record data for the most recent year from the inspection record database 22.

[0035] In step S102, the processing unit 10 performs a predetermined pre-processing by the pre-processing unit 12. For example, the processing unit 10 generates processed sensor data by calculating the moving average and variance of each of the multiple sensor data. In the pre-processing, the calculation results of the moving average and the like at the time of the previous learning may be used. In addition, the processing unit 10 regards the inspection record data, which is the inspection result once a day, as an inspection record from moment to moment, and generates processed inspection record data by associating it with the processed sensor data every 10 minutes. In this way, the processed sensor data and the processed inspection record data can be associated with each date and time as a set.

[0036] In step S103, the processing unit 10 performs an abnormality determination process using the abnormality determination unit 13. That is, the processing unit 10 sets a threshold value for the degree of abnormality of the processed sensor data based on the Hotelling theory, and further determines the presence or absence of an abnormality sign in the water treatment facility 30 (see FIG. 1) based on whether the degree of abnormality is equal to or greater than the threshold value.

[0037] In step S104, the processing unit 10 determines whether or not sensor data related to the abnormality sign has been identified. For example, when the processing unit 10 has previously learned a predetermined classification model and is now updating the classification model, the sensor data related to the abnormality sign of the water treatment facility 30 (see FIG. 1) has already been identified during the past learning. In this way, if the sensor data related to the abnormality sign has already been identified (S104: Yes), the processing of the processing unit 10 proceeds to step S106. On the other hand, if the sensor data related to the abnormality sign has not been identified (S104: No), the processing of the processing unit 10 proceeds to step S105.

[0038] In step S105, the processing unit 10 selects input data by the input data selection unit 14. First, the processing unit 10 acquires an anomaly determination data set including the processed sensor data at each time, the degree of anomaly, and the determination result of the presence or absence of an anomaly from the anomaly determination database 25. Then, the processing unit 10 identifies a combination of processed sensor data that has the highest degree of association with the determination result of the presence or absence of an anomaly based on the Stepwise method, and identifies the processed inspection record data corresponding to this combination. Among the many sensors (not shown) installed in the water treatment facility 30 (see FIG. 1), there are some whose detection values ​​have little relationship to the presence or absence of an anomaly sign in the water treatment facility 30 (i.e., low degree of association), so the detection values ​​of these sensors are excluded from the learning target.

[0039] In step S106, the processing unit 10 performs classification model processing using the classification model unit 15. That is, the processing unit 10 learns a predetermined classification model using the input data (anomaly determination data set) selected in step S105 as teacher data.

[0040] In step S107, the processing unit 10 judges whether the combination of sensor data used to train the classification model is appropriate. First, the processing unit 10 applies predetermined processed sensor data to the trained classification model to judge the presence or absence of an abnormality sign in the water treatment facility 30 (see FIG. 1). Then, the processing unit 10 calculates the false alarm rate and the failure to report rate in judging the presence or absence of an abnormality sign.

[0041] Here, the "false alarm rate" refers to the proportion of cases where the processing unit 10 judges that "there is a sign of abnormality" when the water treatment equipment 30 is actually normal. Also, the "failure to report rate" refers to the proportion of cases where the processing unit 10 judges that "there is no sign of abnormality" when there is actually an abnormality (or a sign of abnormality) in the water treatment equipment 30. For example, an abnormality judgment data set based on inspection records is used as a standard (correct answer) when calculating these false alarm rates and failure to report rates.

[0042] When a predetermined classification model is used, if the false alarm rate is equal to or greater than the first predetermined value, or if the failure rate is equal to or greater than the second predetermined value, the processing unit 10 determines that the combination of sensor data used to learn the classification model is not appropriate (S107: No), and proceeds to the process of step S108. The above-mentioned first predetermined value is a threshold value of the false alarm rate that is a criterion for determining whether or not the combination of sensor data is appropriate in learning the classification model (i.e., whether or not the accuracy of the abnormality sign detection using the classification model falls within an acceptable range), and is set in advance. The second predetermined value is a threshold value of the failure rate that is a criterion for determining whether or not the combination of sensor data is appropriate in learning the classification model, and is set in advance. The magnitude relationship between these "first predetermined value" and "second predetermined value" is not particularly limited.

[0043] When the processing unit 10 calculates the false alarm rate and the failure to report rate, a predetermined mixing matrix based on the anomaly determination data set may be used. Here, the mixing matrix is ​​a matrix in which the actual state of the water treatment facility 30 (see FIG. 1) and the determination results of the processing unit 10 are arranged in a matrix shape. For example, the actual state (normal or abnormal) of the water treatment facility 30 may correspond to each row of the mixing matrix, and the determination results of the processing unit 10 may correspond to each column of the mixing matrix.

[0044] In step S108, the processing unit 10 changes the combination of sensor data. That is, the processing unit 10 selects another combination (a combination different from the one selected in S105) that is affecting the result of the inspection record from among a plurality of combinations of sensor data. After performing the process of step S108, the processing unit 10 returns to the process of step S106, and learns the classification model again.

[0045] In step S107, when the false alarm rate is less than the first predetermined value and the failure rate is less than the second predetermined value, the processing unit 10 judges that the combination of sensor data used to learn the classification model is appropriate (S107: Yes), and ends a series of processes related to learning (END). Note that, when the false alarm rate is less than the first predetermined value and the failure rate is less than the second predetermined value, the first predetermined value and the second predetermined value are set in advance as thresholds so that the abnormality sign detection using the classification model has sufficient accuracy. In this way, the processing unit 10 judges whether the combination is appropriate or not based on the false alarm rate and the failure rate when the presence or absence of an abnormality sign of the water treatment facility 30 (target facility: see FIG. 1) is judged using the processed sensor data of the combination specified by the Stepwise method. Next, the details of each process of steps S102, S103, S105, and S106 shown in FIG. 5 will be described in sequence.

[0046] FIG. 6 is a flowchart of the preprocessing (also see FIG. 2 as appropriate). The details of the process in step S102 in FIG. 5 are shown in the flowchart of FIG. In step S1021, the processing unit 10 accepts a selection input of a processing method for the input data. In the processing of step S1021, for example, a sampling period for associating the sensor data and the inspection record data included in the input data with each date and time is appropriately set by an input operation of the user. In addition, for the sensor data, a calculation method such as a moving average is appropriately set.

[0047] Next, in step S1022, the processing unit 10 generates processed data (processed sensor data and processed inspection record data). For example, the processing unit 10 generates the processed sensor data by calculating a moving average of the sensor data. The processing unit 10 may generate the processed sensor data by dividing the deviation of the sensor data with respect to the moving average by the standard deviation (i.e., standardizing). By performing such standardization, the influence of sensor data with different units on the abnormality sign can be correctly evaluated.

[0048] Furthermore, as the process of step S1022, the processing unit 10 generates processed inspection record data by associating the inspection record data with a predetermined date and time. Usually, the period when a predetermined state quantity is detected by each sensor (not shown) is different from the inspection period by an inspector. In the examples of Fig. 3 and Fig. 4, sensor data such as the treatment flow rate, suction pressure, and transmembrane pressure in the water treatment facility 30 is detected every 10 minutes, whereas inspection by an inspector is performed once a day. Therefore, the processing unit 10 generates processed inspection record data by associating the inspection record data with each date and time of the processed sensor data.

[0049] Next, in step S1023, the processing unit 10 stores the processed data. That is, the processing unit 10 stores the processed sensor data in the processed sensor database 23, and stores the processed inspection record data in the processed inspection record database 24.

[0050] FIG. 7 is an explanatory diagram showing an example of the processed sensor database 23. As shown in FIG. In the example of Figure 7, processed sensor data of treatment flow rate, suction pressure, and transmembrane pressure in water treatment equipment 30 (see Figure 1) is stored in processed sensor database 23 in association with a specified date and time (sampling interval of every 10 minutes).

[0051] FIG. 8 is an explanatory diagram showing an example of the processed inspection record database 24. As shown in FIG. In the example of FIG. 8, as inspection record data indicating the state of the water treatment facility 30 (see FIG. 1), "abnormal" is associated with 0:00 on January 1, and also with 0:10 on January 1. That is, inspection is performed once a day on the water treatment facility 30 (see FIG. 1), and processed inspection record data is associated with each date and time of the processed sensor data (see FIG. 7). For example, if the result of an inspection performed once a day by an inspector is "abnormal," "abnormal" is associated at 10-minute intervals from 0:00 to 23:50 as the processed inspection record data for that day. The same applies when the inspection result is "normal."

[0052] FIG. 9 is a flowchart relating to the abnormality detection process (also see FIG. 2 as appropriate). The details of the process in step S103 in FIG. 5 are shown in the flowchart of FIG. In step S1031, the processing unit 10 acquires processed data (i.e., processed sensor data and processed inspection record data). For example, the processing unit 10 acquires processed sensor data for the most recent year from the processed sensor database 23, and acquires processed inspection record data for the most recent year from the processed inspection record database 24.

[0053] In step S1032, the processing unit 10 applies the Hotelling theory. To specifically explain the Hotelling theory, the processing unit 10 first calculates the degree of abnormality a(x) based on the following formula (1). The degree of abnormality a(x) is a numerical value indicating the degree of abnormality of the water treatment facility 30 (see FIG. 1). Note that x included in formula (1) is the value of the processed sensor data, σ is the variance of the processed sensor data, and μ is the population mean of the processed sensor data.

[0054]

number

[0055] In addition, the abnormality determination unit 13 determines a threshold value a of the abnormality degree a by using the following formula (2) based on the probability density function of the chi-square distribution. thIn addition, α in formula (2) is the occurrence probability of anomaly degree a. Also, χ 2 The part (a|1,1) indicates that the distribution of the degree of anomaly a follows a chi-square distribution with a degree of freedom of 1 and a scale factor of 1. In this way, the processing unit 10 derives the statistical distribution of the processed sensor data based on the Hotelling theory.

[0056]

number

[0057] Next, in step S1033, the processing unit 10 determines whether or not there is a sign of abnormality in the water treatment facility 30. First, the processing unit 10 calculates the degree of abnormality a based on the processed sensor data. Then, the processing unit 10 calculates the degree of abnormality a based on the processed sensor data. th If the abnormality degree a is equal to or greater than the threshold a, the processing unit 10 determines that there is a sign of an abnormality in the water treatment facility 30. th If it is less than this, it is determined that there is no sign of an abnormality in the water treatment facility 30 (that is, the water treatment facility 30 is normal).

[0058] Next, in step S1034, the processing unit 10 combines the data in a predetermined manner. That is, the processing unit 10 associates the processed sensor data and the processed inspection record data at each date and time with an abnormality degree.

[0059] In step S1035, the processing unit 10 applies the inspection record data to normal days, and applies the result of the abnormality determination to abnormal days. Here, a "normal day" is a date when the result of the inspection record of the water treatment equipment 30 (see FIG. 1) was normal. Also, an "abnormal day" is a date when the result of the inspection record of the water treatment equipment 30 was abnormal.

[0060] Inspection of the water treatment equipment 30 (see FIG. 1) by an inspector once a day is actually performed during a certain time period. For example, if the result of the inspection record is "abnormal," it is possible that the water treatment equipment 30 was abnormal during the time period when the inspector performed the inspection, but the water treatment equipment 30 was normal during other time periods. Therefore, using the processed sensor data from moment to moment on the date on which the result of the inspection record was "abnormal," the processing unit 10 determines whether or not there is a sign of abnormality during each time period on that date based on the Hotelling theory, and stores the determination result in the abnormality determination database 25 in association with the processed sensor data.

[0061] Furthermore, when an abnormality actually occurs in the water treatment equipment 30 (see FIG. 1), it rarely returns to normal naturally without maintenance, and the abnormal state often continues. Therefore, on the date on which an experienced inspector judges the water treatment equipment 30 to be "normal", it is highly unlikely that an abnormality occurred in the water treatment equipment 30 before the timing of the inspection. Furthermore, it is also low that an abnormality suddenly occurs on the same day after an experienced inspector judges the water treatment equipment 30 to be "normal". Therefore, in this embodiment, the inspector's judgment is given priority for dates on which the inspection record indicates "normal", and the processing unit 10 regards the moment-to-moment state of the water treatment equipment 30 (see FIG. 1) on those dates as being "normal".

[0062] To specifically explain the process of step S1035, the processing unit 10 associates "abnormal" (with signs of abnormality) with the processed sensor data for the date on which the result of the inspection record was "abnormal" when the degree of abnormality was equal to or greater than a predetermined threshold, and associates "normal" (without signs of abnormality) with the processed sensor data for the date on which the result of the inspection record was "abnormal" when the degree of abnormality was less than the predetermined threshold. This makes it possible to obtain a more accurate learning result (learning result of a classification model described later) when the result of the inspection record once a day is "abnormal" than if the "abnormal" state had continued uniformly for the entire 24 hours of that day.

[0063] Furthermore, the processing unit 10 regards the processed sensor data on a date where the result of the inspection record was "normal" as being "normal" at each time on that day. In other words, even if there is processed sensor data with an abnormality level equal to or greater than a predetermined threshold, the processing unit 10 prioritizes the judgment of an experienced inspector and treats the data as "normal" since it is highly likely to be a transient change associated with the operation / stop of the device.

[0064] FIG. 10A is an explanatory diagram regarding the process when the result of the inspection record indicates an abnormality. The upper part of Fig. 10A shows the transition of the degree of abnormality of the water treatment equipment 30 (see Fig. 1) on a specified date, and the middle part shows the inspection record (processed inspection record data). The lower part of Fig. 10A shows the judgment result by the abnormality judgment unit 13. The horizontal axis of each of the upper, middle, and lower parts of Fig. 10A represents time. It is assumed that the inspection by the inspector was performed between times t2 and t3, and the result of the inspection record for this day was "abnormal."

[0065] Referring to the transition of the abnormality degree in the upper part of FIG. 10A, the abnormality degree is the threshold a th The anomaly level was greater than or equal to the threshold a after time t1. th In the processed inspection record data in the middle of FIG. 10A, when the result of the inspection record once a day is "abnormal", the moment-to-moment condition of the water treatment facility 30 (see FIG. 1) on that day is deemed to be "abnormal". On the other hand, as shown in the lower part of FIG. 10A, the result of the abnormality determination based on the Hotelling theory indicates that the facility was normal up to time t1, but became abnormal (signs of abnormality were present) after time t1. In this way, for dates on which the result of the inspection record is "abnormal", the processing unit 10 determines the presence or absence of an abnormality at each time based on the Hotelling theory. This makes it possible to use data reflecting the actual condition of the water treatment facility 30 as training data for the classification model.

[0066] FIG. 10B is an explanatory diagram regarding the process when the result of the inspection record is normal. The vertical and horizontal axes of the upper, middle, and lower time charts in FIG. 10B are the same as those in FIG. 10A. th The abnormality level was greater than or equal to the threshold a during the other time periods. th It is less than that.

[0067] In the processed inspection record data in the middle part of FIG. 10B, if the result of the inspection record once a day is "normal", the moment-to-moment condition of the water treatment facility 30 on that day is considered to be "normal". Also, as shown in the bottom part of FIG. 10B, the inspection record data based on the processing result of the abnormality determination unit 13 also considers the moment-to-moment condition to be "normal". In other words, if the degree of abnormality temporarily exceeds the threshold value a th Even if there is a time period (time t11 to t12) during which the above condition occurs, if the inspection record is "normal", the inspector's judgment is given priority, and the water treatment equipment 30 (see Figure 1) is deemed to have been "normal" during that time period.

[0068] In this way, for a date during a predetermined past period when an inspector's inspection of the water treatment facility 30 (target facility: see FIG. 1) was normal, the processing unit 10 generates anomaly determination data by assuming that the water treatment facility 30 was normal at each time on that date. Also, for a date when the inspection was abnormal, the processing unit 10 generates anomaly determination data by associating the presence or absence of anomaly signs based on the statistical distribution of processed sensor data with each time on that date.

[0069] FIG. 11 is an explanatory diagram showing an example of the abnormality determination database 25. As shown in FIG. In the example of FIG. 11, the transmembrane pressure value included in the "sensor data" (processed sensor data), the degree of abnormality, and the presence or absence of an abnormality (i.e., the presence or absence of an abnormality sign) are associated with each other. The value of the degree of abnormality is "1.63", which is the threshold value a thSince the measured value is less than or equal to the measured value, it is determined to be "normal." In the processed inspection record database 24 of FIG. 8, the date is uniformly determined to be "abnormal" based on the inspection record for January 1, but in the abnormality determination database 25 shown in FIG. 11, it is determined to be "normal" at 0:00 and 0:10 on January 1 based on the Hotelling theory. Although omitted in FIG. 11, other processed sensor data such as the processed flow rate and suction pressure are also included in the abnormality determination database 25.

[0070] FIG. 12 is a flowchart relating to the selection of input data (also see FIG. 2 as appropriate). The details of the process in step S105 in FIG. 5 are shown in the flowchart of FIG. In step S1051, the processing unit 10 acquires a data set of the determination result of the presence or absence of an abnormality. That is, the processing unit 10 acquires the processed sensor data, the degree of abnormality, and the determination result of the presence or absence of an abnormality (presence or absence of an abnormality sign) at each time from the abnormality determination database 25 (see FIG. 11) by the input data selection unit 14.

[0071] In step S1052, the processing unit 10 applies the Stepwise method. Specifically, the processing unit 10 performs regression analysis based on one or more types of processed sensor data (explanatory variables) selected from the anomaly determination database 25 (see FIG. 11). Then, the processing unit 10 calculates the degree of association between each combination of processed sensor data and the determination result (objective function) of the presence or absence of an anomaly sign based on the Hotelling theory. For example, the processing unit 10 selects the processing flow rate from three processed sensor data, the processing flow rate, the suction pressure, and the transmembrane pressure, and calculates the degree of association between the result of the presence or absence of an anomaly in the anomaly determination database 25 (objective variable) and the processing flow rate (explanatory variable). The same process is performed for all combinations when selecting one, two, or three types from a total of three types, the processing flow rate, the suction pressure, and the transmembrane pressure.

[0072] In step S1053, the processing unit 10 identifies a combination of sensor data with high accuracy in detecting an abnormality sign. That is, the processing unit 10 identifies a combination of types of processed sensor data with the highest degree of association with the determination result (objective function) of the presence or absence of an abnormality sign based on the Hotelling's theory.

[0073] In this way, the processing unit 10 uses the presence or absence of an abnormality sign of the water treatment facility 30 (target facility: see FIG. 1) in the abnormality determination data as a response variable and a plurality of processed sensor data as explanatory variables, and identifies a predetermined combination based on the degree of association with the objective variable from among the explanatory variables based on the stepwise method. Then, the combination of processed sensor data identified in this way is used by the processing unit 10 when detecting an abnormality sign of the water treatment facility 30 (target facility: see FIG. 1) after learning the classification model.

[0074] FIG. 13 is a flowchart relating to classification model processing (also see FIG. 2 if appropriate). The details of the process in step S106 in FIG. 5 are shown in the flowchart of FIG. In step S1061, the processing unit 10 acquires a data set with high accuracy in detecting a sign of anomaly. That is, the processing unit 10 acquires a data set corresponding to the combination of types of processed sensor data identified by the Stepwise method (S1052, S1053 in FIG. 12). These data sets are used as teacher data when the processing unit 10 learns a predetermined classification model.

[0075] In step S1062, the processing unit 10 performs machine learning by applying random forest. In random forest, first, data is randomly restored and extracted from the original data (the data set acquired in S1061) to create multiple data groups. Then, a decision tree is created for each of the multiple data groups, and the presence or absence of signs of abnormality in the water treatment facility 30 is determined individually based on these decision trees. Furthermore, a final determination result is issued based on a majority vote of the multiple data groups. By learning a classification model based on such random forest, the accuracy of detecting signs of abnormality in the water treatment facility 30 (see FIG. 1) is improved. The learning results of the classification model are stored in the classification model database 26.

[0076] In step S1063, the processing unit 10 accepts a setting input of a learning period and an inference period. The "learning period" is a period specified when a user acquires a data set used as teacher data for learning a predetermined classification model. For example, the most recent one year may be set as the learning period. The "inference period" is a period specified by a user when inference (determination of the presence or absence of an abnormality sign) is performed using a classification model. For example, a date on which sensor data has been obtained but an inspection by an inspector has not yet been performed may be specified as the inference period. The process of step S1063 may be performed before step S1062.

[0077] In step S1064, the processing unit 10 determines whether or not there is a sign of abnormality in the water treatment facility 30. That is, the processing unit 10 determines whether or not there is a sign of abnormality in the water treatment facility 30 (target facility) by performing random forest using the processed sensor data of the combination identified based on the Stepwise method.

[0078] FIG. 14 is a flowchart of processing related to inference by the anomaly sign detection system (also see FIG. 2 as appropriate). In step S201, the processing unit 10 reads uninspected sensor data for a predetermined date. In step S202, the processing unit 10 performs a predetermined pre-processing by the pre-processing unit 12. The processing of step S202 is the same as step S102 in Fig. 5. That is, the processing unit 10 performs a predetermined pre-processing on the sensor data to generate processed sensor data. Also, the processing unit 10 performs a predetermined pre-processing on the inspection record data to generate processed inspection record data.

[0079] Next, in step S203, the processing unit 10 performs inference based on the classification model using the classification model unit 15 (detection process). That is, the processing unit 10 determines the presence or absence of a sign of abnormality in the water treatment facility 30 based on the processed sensor data acquired on an uninspected date when the water treatment facility 30 (target facility) has not yet been inspected, and the classification model. Specifically, the processing unit 10 reads the classification model from the classification model database 26, and applies random forest to the processed sensor data to determine the presence or absence of a sign of abnormality in the water treatment facility 30. The processing result based on the classification model is stored in the detection result database 27.

[0080] In step S204, the processing unit 10 displays the processing result. That is, the processing unit 10 displays on a display device (not shown) whether or not there are signs of abnormality in the water treatment equipment 30 for a specified date on which an inspector has not yet performed an inspection. This allows the inspector, etc. to take measures such as maintenance of the water treatment equipment 30 early if there are signs of abnormality. After performing the processing of step S204, the processing unit 10 ends a series of processes related to inference (END).

[0081] Note that inference based on the classification model may be repeated at a predetermined cycle (for example, every 10 minutes). Also, while detecting signs of abnormality in the water treatment facility 30 using the abnormality sign detection system 100, an inspector may appropriately inspect the water treatment facility 30.

[0082] <Effects> In this embodiment, for dates on which the inspector's inspection results are normal, the water treatment facility 30 is assumed to be normal at each time, and abnormality determination data is generated. On the other hand, for dates on which the inspection results are abnormal, the presence or absence of abnormality signs is determined based on the statistical distribution of the processed sensor data, and abnormality determination data is generated. In this way, the way in which the teacher data is reflected varies depending on the results of the inspector's inspection records. This improves the accuracy of the abnormality sign detection system 100 when detecting abnormality signs in the water treatment facility 30. As a result, the presence or absence of abnormality signs can be detected with high accuracy, close to that of an experienced inspector, based on the sensor data and the inspection records taken once a day.

[0083] Also, there is no particular need for the inspector to increase the frequency of inspecting the water treatment equipment 30. For example, the inspector can inspect the water treatment equipment 30 once a day, while outputting the results of whether or not there are any signs of abnormality in the water treatment equipment 30 at predetermined time intervals (for example, every 10 minutes). Therefore, the workload of the inspector can be reduced.

[0084] Moreover, according to this embodiment, by using the Hotelling theory, it is possible to appropriately identify outliers from sensor data obtained in a predetermined period. Furthermore, by using the Stepwise method, it is possible to identify sensor data that contributes to high accuracy in detecting signs of abnormality in the water treatment facility 30. Furthermore, by using a random forest when performing learning and inference, it is possible to further improve the accuracy of detecting signs of abnormality.

[0085] <<Variations>> The abnormality sign detection system 100 and the abnormality sign detection method according to the present disclosure have been described in the above embodiment, but are not limited to this description and can be modified in various ways. For example, in the embodiment, the case where the water treatment facility 30 is considered to be normal at each time on the date on which the inspection result was "normal" has been described, but is not limited to this. That is, the processing unit 10 may consider the date on which the inspection result was "normal" to be normal until the end of the inspection time period, and perform an abnormality determination based on the Hotelling theory for each time after the end of the inspection time period. With such processing, the same effect as in the embodiment can be achieved, and an abnormality sign can be appropriately detected even in cases where an abnormality sign occurs after an inspector has determined that the condition is normal.

[0086] In the embodiment, the abnormality determination unit 13 determines whether or not there is an abnormality sign based on the Hotelling's theory, but the present invention is not limited to this. For example, the abnormality determination unit 13 may determine whether or not there is an abnormality sign based on the k-nearest neighbor method. In the embodiment, the input data selection unit 14 selects the input data based on the Stepwise method, but the present invention is not limited to this. For example, the input data may be selected based on the forced input method. In the embodiment, a case has been described in which a random forest is applied to the processing of the classification model unit 15, but this is not limiting. For example, a support vector machine or a clustering technique may be used.

[0087] In the embodiment, for a date on which an inspection result of the target equipment by an inspector was abnormal in a predetermined period in the past, the processing unit 10 generates anomaly determination data by associating the presence or absence of anomaly signs based on the statistical distribution of the processed sensor data with each time on the date, but this is not limited to the above. For example, for a date on which an inspection result of the target equipment was abnormal, the processing unit 10 may determine the presence or absence of anomaly signs by comparing each measurement value of the processed sensor data with a predetermined threshold. The threshold may be a fixed value or may be changed as appropriate. In addition, for a date on which an inspection result was abnormal, the processing unit 10 may determine the presence or absence of anomaly signs based on clustering.

[0088] Furthermore, the processing of the abnormality sign detection system 100 (the abnormality sign detection method) may be executed as a predetermined program of a computer. The program can be provided via a communication line, or can be written to a predetermined recording medium and distributed.

[0089] In addition, the present disclosure is not limited to the embodiments, and includes various modified examples. For example, the embodiments are described in detail to clearly explain the present disclosure, and the present disclosure is not necessarily limited to those including all of the configurations described. In addition, the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. In addition, the above-mentioned configurations, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as the program, table, file, etc. that realizes each function can be stored in a memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD. In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0090] 10 Processing section 11 Input data acquisition section 12 Pretreatment section 13 Abnormality determination section 14 Input data selection section 15 Classification Model Section 21 Sensor Database 22 Inspection record database 23 Processed Sensor Database 24 Processed inspection records database 25 Anomaly Judgment Database 26 Classification Model Database 27 Detection Results Database 30 Water treatment equipment (target equipment) 40 PLC 50 Computer 100 Anomaly Prediction Detection System S203 Step (detection process)

Claims

1. a processing unit that detects a sign of an abnormality in the target equipment based on a classification model using as training data a data set including processed sensor data obtained by performing a predetermined preprocessing on detection values ​​of a plurality of sensors in the target equipment and abnormality determination data indicating the presence or absence of an abnormality sign in the target equipment, The processing unit generates the anomaly determination data by assuming that the target equipment was normal at each time on any date during a specified period in the past when the inspection results of the target equipment by an inspector were normal.

2. The processing unit generates the abnormality determination data by associating the presence or absence of an abnormality sign based on a statistical distribution of the processed sensor data with each time on the date on which the inspection result is abnormal.

2. The abnormality sign detection system according to claim 1,

3. The processing unit derives the statistical distribution based on the Hotelling theory.

3. The abnormality sign detection system according to claim 2,

4. the processing unit uses the presence or absence of an abnormality sign of the target equipment in the abnormality determination data as an objective variable, and the plurality of processed sensor data as explanatory variables, and identifies a predetermined combination based on a degree of association between the explanatory variables and the objective variable from among the plurality of explanatory variables based on a stepwise method, and uses the processed sensor data of the combination to detect an abnormality sign of the target equipment after learning the classification model.

4. The abnormality sign detection system according to claim 3,

5. The processing unit performs a random forest using the processed sensor data of the combination identified based on the Stepwise method to determine whether or not there is a sign of an abnormality in the target equipment.

5. The abnormality sign detection system according to claim 4,

6. The processing unit determines whether the combination is appropriate based on a false alarm rate and a false alarm rate when determining the presence or absence of an abnormality sign in the target equipment using the processed sensor data of the combination identified by the Stepwise method.

5. The abnormality sign detection system according to claim 4,

7. The processing unit determines whether or not there is a sign of an abnormality in the target equipment based on the processed sensor data acquired on an uninspected date on which the target equipment has not yet been inspected and the classification model.

2. The abnormality sign detection system according to claim 1,

8. A detection process for detecting a sign of an abnormality in the target equipment based on a classification model using as training data a data set including processed sensor data obtained by performing a predetermined pre-processing on detection values ​​of a plurality of sensors of the target equipment and abnormality determination data indicating the presence or absence of an abnormality sign in the target equipment, In the detection process, for any date during a specified period in the past when an inspector's inspection results of the target equipment were normal, the target equipment was deemed to have been normal at each time on that date, and the abnormality determination data is generated based on this detection process.

Images