State determination device, state determination program, and state determination system

Abstract

To provide a state determination device, a state determination program, and a state determination system capable of constructing an anomaly detection model with high accuracy in determining whether it is normal or abnormal through learning processing, even when time-series data that does not include an abnormal occurrence section is obtained.SOLUTION: A state determination device (18) calculates a modulation pattern PT indicating time changes of relative intensity of vibration in an abnormal occurrence section using abnormal data D2 including an abnormal occurrence section, generates simulated abnormal data D3 simulating the abnormal data D2 by multiplying each sample value of a physical quantity belonging to a processing section extracted from all time-series sections by a modulation pattern PT regarding processing object data (D1), and performs learning processing on an anomaly detection model (M) using the simulated abnormal data D3.SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] The present disclosure relates to a state determination device, a state determination program, and a state determination system. [Background technology]

[0002] BACKGROUND ART A condition determination system has been known in the past that sequentially measures a physical quantity correlated with vibrations of a monitored object, performs analytical processing on the obtained time-series data, and determines whether the monitored object is in a normal state or an abnormal state.

[0003] Patent Document 1 discloses a method of calculating an error for each frequency of diagnostic data and estimating the cause of an abnormality in a target facility based on an evaluation of the calculated error.

[0004] Patent Document 2 discloses a device that selects a threshold value for determining whether vibration data during operation is normal or abnormal based on the verification results regarding the appropriateness of abnormality detection sensitivity.

[0005] Patent Document 3 discloses a method of calculating a plurality of dimensional vibration parameters that characterize the vibration of a rotating machine, and determining whether the rotating machine is good or bad based on the analysis results of the vibration parameters using principal component analysis. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2021-144054 [Patent Document 2] Japanese Patent Application Publication No. 2019-027788 [Patent Document 3] Japanese Patent Application Laid-Open No. 2008-058191 Summary of the Invention [Problem to be solved by the invention]

[0007] Generally speaking, this type of anomaly detection model often uses "unsupervised learning," a method of training without providing correct answers to the training data. The reasons for this are: [1] it is difficult to collect a large amount of data because abnormal states are expected to occur infrequently, [2] abnormal states are diverse, and unknown anomalies may not be detected well, or [3] it is difficult to strictly define normal and abnormal.

[0008] However, Patent Documents 1 to 3 do not take into account the case where, when constructing an anomaly detection model through a learning process using time series data, the entire time series does not include a section indicating that an abnormality has occurred in the condition of the monitored object (hereinafter referred to as the abnormality occurrence section).

[0009] The present disclosure has been made in consideration of the above-mentioned problems, and its purpose is to provide a state determination device, a state determination program, and a state determination system that are capable of constructing an anomaly detection model that has high accuracy in determining whether something is normal or abnormal through a learning process, even when time-series data that does not include an abnormality occurrence section is obtained. [Means for solving the problem]

[0010] In order to solve the above problem, a state determination device according to one aspect of the present disclosure includes: a data acquisition unit that acquires abnormality data, which is time-series data of a physical quantity correlated with vibration of a monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the state of the monitored object within an entire section of the time series; and processing target data, which is time-series data of the physical quantity and is an object to be processed; a pattern calculation unit that uses the abnormality data acquired by the data acquisition unit to calculate a modulation pattern indicating a change over time in the relative intensity of the vibration within the abnormality occurrence section; and a data generation unit that extracts, from all sections of the time series of the processing target data acquired by the data acquisition unit, a processing section equal to the length of the abnormality occurrence section, and multiplies each sample value of the physical quantity belonging to the processing section by the modulation pattern to generate simulated abnormality data that simulates the abnormal data.

[0011] A state determination program according to another aspect of the present disclosure causes one or more computers to execute the following steps: an acquisition step of acquiring abnormality data, which is time-series data of a physical quantity correlated with vibration of a monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the state of the monitored object within an entire section of the time series; and processing target data, which is time-series data of the physical quantity and is also an object to be processed; a calculation step of calculating, using the acquired abnormality data, a modulation pattern indicating a change over time in the relative intensity of the vibration within the abnormal section; and a generation step of extracting, with respect to the acquired processing target data, a processing section equal to the length of the abnormality occurrence section from all sections of the time series, and multiplying each sample value of the physical quantity belonging to the processing section by the modulation pattern to generate simulated abnormality data that simulates the abnormal data.

[0012] According to another aspect of the present disclosure, a status determination system includes a monitored object exhibiting behavior accompanied by vibration; a sensor that measures a physical quantity correlated with the vibration of the monitored object and outputs a measurement signal; and a status determination device that acquires the measurement signal from the sensor and outputs a determination result as to whether the status of the monitored object is normal or abnormal. The status determination device includes a data acquisition unit that acquires abnormality data, which is time-series data of the physical quantities correlated with the vibration of the monitored object and includes an abnormality occurrence section that indicates that an abnormality has occurred in the status of the monitored object within an entire section of the time series, and processing target data, which is time-series data of the physical quantities and is to be processed; a pattern calculation unit that uses the abnormality data acquired by the data acquisition unit to calculate a modulation pattern that indicates a change over time in the relative intensity of the vibration within the abnormality occurrence section; and a data generation unit that extracts, with respect to the processing target data acquired by the data acquisition unit, a processing section that is equal to the length of the abnormality occurrence section from all sections of the time series, and generates simulated abnormality data that simulates the abnormal data by multiplying each sample value of the physical quantity belonging to the processing section by the modulation pattern. [Effects of the Invention]

[0013] According to the present disclosure, even when time-series data that does not include an abnormality occurrence section is obtained, an anomaly detection model that has high accuracy in determining whether something is normal or abnormal can be constructed through learning processing. [Brief explanation of the drawings]

[0014] [Figure 1] 1 is an overall configuration diagram of a state determination system according to an embodiment of the present disclosure. [Figure 2] 2 is a block diagram showing an example of the configuration of the state determination device in FIG. 1. FIG. [Figure 3] FIG. 10 is a diagram illustrating an example of a method for generating time-series feature amounts. [Figure 4] FIG. 3 is a functional block diagram relating to the anomaly detection model shown in FIG. 2. [Figure 5] FIG. 2 is a sequence diagram relating to the operation of the state determination system shown in FIG. [Figure 6] FIG. 10 is a diagram illustrating an example of a determination result of abnormal data by an anomaly detection model. [Figure 7] FIG. 3 is a functional block diagram of a data generating unit shown in FIG. 2. [Figure 8] 10 is a detailed flowchart of a method for generating simulated abnormality data (steps SP14 and SP24 in FIG. 5). [Figure 9] FIG. 10 is a diagram illustrating an example of a calculation result of a modulation pattern. [Figure 10] FIG. 10 is a diagram showing an example of a processing result of normal data. [Figure 11] FIG. 3 is a functional block diagram of a model evaluation unit shown in FIG. 2. [Figure 12] 10 is a detailed flowchart illustrating a method for evaluating an anomaly detection model in provisional learning (step SP16 in FIG. 5). [Figure 13] FIG. 10 is a diagram illustrating an example of a method for calculating the number of erroneous determinations. [Figure 14] FIG. 10 is a diagram illustrating an example of a method for calculating a grace period. [Figure 15] FIG. 10 is a diagram schematically illustrating an example of a method for determining model parameters. DETAILED DESCRIPTION OF THE INVENTION

[0015] First, some aspects of the disclosure will be described.

[0016] A condition determination device according to a first aspect of the present disclosure includes: a data acquisition unit that acquires abnormality data, which is time-series data of physical quantities correlated with vibrations of a monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the condition of the monitored object within an entire section of the time series; and processing target data, which is time-series data of the physical quantities and is an object to be processed; a pattern calculation unit that uses the abnormality data acquired by the data acquisition unit to calculate a modulation pattern indicating a change over time in the relative intensity of the vibrations within the abnormality occurrence section; and a data generation unit that extracts, from all sections of the time series of the processing target data acquired by the data acquisition unit, a processing section equal to the length of the abnormality occurrence section, and multiplies each sample value of the physical quantities belonging to the processing section by the modulation pattern to generate simulated abnormality data that simulates the abnormal data.

[0017] In the state determination device according to the second aspect of the present disclosure, the modulation pattern may be obtained by dividing each sample value of the physical quantity belonging to the abnormality occurrence section by a representative value of the physical quantity belonging to a normal section of all sections of a time series, excluding the abnormality occurrence section.

[0018] In the state determination device according to the third aspect of the present disclosure, the abnormality occurrence section may include a transition section during which the state of the monitored object transitions from normal to abnormal.

[0019] In the state determination device according to the fourth aspect of the present disclosure, the processing section may be a section in which the physical quantity is relatively large compared to other sections among all sections of the time series.

[0020] In the state determination device according to the fifth aspect of the present disclosure, the simulated abnormality data may be used for learning or verification processing of an anomaly detection model that receives time-series data of the physical quantity as an input and outputs a determination result as to whether the state of the monitored object is normal or abnormal.

[0021] The state determination device according to a sixth aspect of the present disclosure may further include a model evaluation unit that receives time-series data of the physical quantity as an input, calculates an evaluation index related to the detection performance of an anomaly detection model that outputs a determination result as to whether the state of the monitored object is normal or abnormal, and evaluates the anomaly detection model based on the evaluation index.

[0022] In the state determination device according to a seventh aspect of the present disclosure, the evaluation index may include the number of erroneous determinations, and the number of erroneous determinations may be the number of times that an erroneous determination that an abnormality exists is made within a determination period that does not include the abnormality occurrence section, with respect to a time transition of the determination result obtained by inputting time series data of the physical quantity into the anomaly detection model.

[0023] In the state determination device according to an eighth aspect of the present disclosure, the evaluation index may include a grace period, and the grace period may be, with respect to a time progression of the determination result obtained by inputting time series data of the physical quantity into the anomaly detection model, a period from a point in time when an erroneous determination of an anomaly is first made to a point in time when the anomaly occurs, which is included in the time series data of the physical quantity, within a determination period including the anomaly occurrence section.

[0024] In the state determination device according to a ninth aspect of the present disclosure, the model evaluation unit may calculate the number of erroneous determinations and the grace period corresponding to a plurality of model parameter candidates for identifying a model structure of the anomaly detection model, and select one model parameter candidate from among the plurality of model parameter candidates according to a combination of the number of erroneous determinations and the grace period.

[0025] In the state determination device according to a tenth aspect of the present disclosure, the model evaluation unit may select the one model parameter candidate that makes the number of erroneous determinations below a threshold and that provides the longest grace period.

[0026] The state determination device according to an eleventh aspect of the present disclosure may further include a learning processing unit that performs a learning process on the anomaly detection model using the anomaly data acquired by the data acquisition unit or the simulated anomaly data generated by the data generation unit.

[0027] In a state determination device according to a twelfth aspect of the present disclosure, the monitored object may be a rotating machine installed within a facility, the model evaluation unit may provisionally determine the anomaly detection model for each of the facilities or each of the types of the rotating machines, and the learning processing unit may finally determine the anomaly detection model to be applied to each of the rotating machines by performing additional or re-learning processing on the anomaly detection model provisionally determined by the model evaluation unit.

[0028] In a state determination device according to a thirteenth aspect of the present disclosure, the anomaly detection model determines whether the state of the monitored object is normal or abnormal based on a magnitude relationship between an error index related to a reconstruction error when the time-series feature quantities related to the physical quantities are reconstructed through dimensionality compression and dimensionality restoration, and a threshold value corresponding to the error index, and the monitored object may be a rotating machine, and the physical quantities may include a vibration speed, a vibration acceleration, and a temperature of the rotating machine.

[0029] In a state determination device according to a fourteenth aspect of the present disclosure, the error index may be a linear sum of a first error, which is a standardized or normalized reconstruction error of the vibration velocity, a second error, which is a standardized or normalized reconstruction error of the vibration acceleration, and a third error, which is a standardized or normalized reconstruction error of the temperature, and a weighting coefficient of the third error included in the linear sum may be smaller than the weighting coefficients of the first error and the second error included in the linear sum.

[0030] In the state determination device according to a fifteenth aspect of the present disclosure, the anomaly detection model may include an unsupervised learner that reconstructs time-series feature quantities related to the physical quantities through dimensionality reduction and dimensionality restoration of the time-series feature quantities, and the time-series feature quantities may include a plurality of partial time-series data extracted from the time-series data of the physical quantities by sequentially sliding an extraction interval starting from a reference time point.

[0031] A state determination program according to a sixteenth aspect of the present disclosure causes one or more computers to execute the following steps: an acquisition step of acquiring abnormality data, which is time-series data of physical quantities correlated with vibrations of a monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the state of the monitored object within an entire section of the time series; and processing target data, which is time-series data of the physical quantities and is also an object to be processed; a calculation step of calculating, using the acquired abnormality data, a modulation pattern indicating a change over time in the relative intensity of the vibration within the abnormality occurrence section; and a generation step of extracting, with respect to the acquired processing target data, a processing section equal to the length of the abnormality occurrence section from all sections of the time series, and multiplying each sample value of the physical quantities belonging to the processing section by the modulation pattern to generate simulated abnormality data that simulates the abnormal data.

[0032] a data acquisition unit that acquires abnormality data, which is time-series data of the physical quantities correlated with the vibrations of the monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the state of the monitored object within an entire section of the time series, and processing object data, which is time-series data of the physical quantities and is to be processed; a pattern calculation unit that calculates, using the abnormality data acquired by the data acquisition unit, a modulation pattern indicating a change over time in the relative intensity of the vibrations within the abnormality occurrence section; and a data generation unit that extracts, with respect to the processing object data acquired by the data acquisition unit, a processing section equal to the length of the abnormality occurrence section from all sections of the time series, and generates simulated abnormality data that simulates the abnormal data by multiplying each sample value of the physical quantities belonging to the processing section by the modulation pattern.

[0033] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. To facilitate understanding of the description, identical components and steps in each drawing will be denoted by the same reference numerals whenever possible, and redundant explanations will be omitted. Furthermore, when terms such as "first" and "second" are used in this specification or claims, unless otherwise specified, they do not represent any order or importance, but are used to distinguish one configuration from another. Furthermore, the term "part" may be replaced with other terms such as unit, module, device, or element.

[0034] [Configuration of the state determination system 10] <Overall structure> FIG. 1 is an overall configuration diagram of a condition determination system 10 according to an embodiment of the present disclosure. The condition determination system 10 is provided to determine the condition of a monitored object. In the example of FIG. 1, the monitored object is a rotating machine 14 provided in a facility such as an oil refinery 12. Specifically, the condition determination system 10 is configured to include one or more (two in the example of FIG. 1) rotating machines 14, a process monitoring device 16, and a condition determination device 18. Here, the process monitoring device 16 and the condition determination device 18 are configured to be able to communicate with each other via a network NT.

[0035] The rotating machines 14 are configured, for example, by electric motors, generators, or turbines. Each rotating machine 14 is equipped with sensors 20, 21, and 22 for measuring physical quantities correlated with the vibration of the rotating machine 14. Examples of these physical quantities include [1] physical quantities indicating the operating state, including displacement, speed, acceleration, or jerk, or [2] physical quantities related to the operating environment of the rotating machine 14, including temperature or humidity. In particular, in the case of the rotating machine 14, excessive frictional heat is generated when an abnormality occurs, so abnormalities can also be detected by measuring the temperature.

[0036] The sensor 20 measures the vibration velocity of the rotating machine 14 and outputs a measurement signal indicating a time series of the vibration velocity to the process monitoring device 16. The sensor 21 measures the vibration acceleration of the rotating machine 14 and outputs a measurement signal indicating a time series of the vibration acceleration to the process monitoring device 16. The sensor 22 measures the internal temperature, surface temperature, or ambient temperature of the rotating machine 14 and outputs a measurement signal indicating a time series of the temperature to the process monitoring device 16.

[0037] The process monitoring device 16 is a computer for monitoring the work processes within the refinery 12. The process monitoring device 16 receives, for example, measurement signals from the sensors 20 to 22 and provides the obtained time-series data to the status determination device 18 periodically or irregularly. The process monitoring device 16 also receives the determination results from the status determination device 18 and outputs these determination results to an output device (not shown). This allows the operator to quickly grasp whether there is a sign of an abnormality occurring in the rotating machine 14 or whether an abnormality has actually occurred.

[0038] The state determination device 18 is a server computer that performs overall control regarding state determination of the rotating machine 14, and may be either a cloud-based or on-premise type. Here, the state determination device 18 is illustrated as a single computer, but instead, the state determination device 18 may be a group of computers that form a distributed system.

[0039] <Configuration of state determination device 18> Fig. 2 is a block diagram showing an example of the configuration of the state determination device 18 in Fig. 1. Specifically, the state determination device 18 includes a communication unit 30, a control unit 32, and a storage unit 34.

[0040] The communication unit 30 is an interface for transmitting and receiving electrical signals to and from an external device, which enables the state determination device 18 to acquire measurement data sequentially output from the sensors 20 to 22 from the process monitoring device 16, and to supply the process monitoring device 16 with a determination result regarding the state of the rotating machine 14.

[0041] The control unit 32 is configured by a processor including a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The control unit 32 reads and executes programs and data stored in the storage unit 34, thereby functioning as a data acquisition unit 40, a data generation unit 42, a learning processing unit 44, a model evaluation unit 46, a data determination unit 48, and an output instruction unit 50.

[0042] The data acquisition unit 40 acquires various data related to the learning process or the judgment process, such as the normal data D1, the abnormal data D2, the simulated abnormal data D3, or the judgment target data D4. The data acquisition unit 40 may acquire data via communication from an external device including the process monitoring device 16, or may acquire data by reading it from the storage unit 34.

[0043] The data generation unit 42 performs preprocessing on the data acquired by the data acquisition unit 40 to generate various data to be used in the learning process, the judgment process, or the verification process. The data generation unit 42 also generates simulated abnormal data D3 by combining the normal data D1 and the abnormal data D2. The specific configuration of the data generation unit 42 will be described in detail with reference to FIG. 4.

[0044] The learning processing unit 44 performs a learning process on the anomaly detection model M using the anomaly data D2 or the simulated anomaly data D3. This "anomaly detection model M" refers to a mathematical model that receives time-series data of physical quantities as input and outputs a determination result as to whether the state of the rotating machine 14 is normal or abnormal. The learning processing unit 44 performs [1] "tentative learning" to tentatively determine the anomaly detection model M for each refinery 12 or each type of rotating machine 14, and [2] "final learning" to finally determine the anomaly detection model M to be applied to each individual rotating machine 14 through additional or re-learning processing. Through this learning process, each value of the tentative learning parameter set 56 or each value of the final learning parameter set 58 is determined.

[0045] The model evaluation unit 46 performs evaluation processing to evaluate the detection performance of the anomaly detection model M constructed through the learning processing of the learning processing unit 44. This evaluation processing includes, for example, [1] a "calculation operation" to calculate evaluation indicators related to the detection performance of the anomaly detection model M (for example, the number of erroneous determinations and the grace period), or [2] a "determination operation" to determine model parameters for identifying the anomaly detection model M. The specific configuration of the model evaluation unit 46 will be described in detail with reference to FIG. 11.

[0046] The data determination unit 48 uses the anomaly detection model M specified by the learning parameter group 58 to perform a determination process as to whether the rotating machine 14 is normal or abnormal with respect to the determination target data D4 acquired by the data acquisition unit 40. The data determination unit 48 may, for example, determine whether the rotating machine 14 is normal or abnormal each time the data acquisition unit 40 acquires the determination target data D4, or may periodically or irregularly determine whether the rotating machine 14 is normal or abnormal based on the determination target data D4 accumulated by the data acquisition unit 40. The specific configuration of the anomaly detection model M will be described in detail with reference to FIG. 4.

[0047] The output instruction unit 50 instructs the output device to present the judgment result obtained by the data judgment unit 48 to the operator. This "output device" may be a device provided in an external device including the process monitoring device 16, or may be an output device (not shown) provided in the status judgment device 18. The output may be in any form that stimulates the five senses, such as sight, hearing, and touch.

[0048] The storage unit 34 stores programs and data necessary for the control unit 32 to control each component. The storage unit 34 is non-transitory and is composed of a computer-readable storage medium. Here, the computer-readable storage medium is composed of [1] a storage device such as a hard disk drive (HDD) or solid state drive (SSD) built into the computer system, or [2] a portable medium such as a magneto-optical disk, a read-only memory (ROM), a compact disc (CD)-ROM, or a flash memory.

[0049] In the example of FIG. 2, the storage unit 34 stores a provisional learning data group 52, a formal learning data group 54, a provisional learning parameter group 56, a formal learning parameter group 58, judgment target data D4, and judgment result information 60.

[0050] The tentative learning data group 52 is a collection of learning data used for tentative learning. Here, "tentative learning" means learning performed on a group basis, for example, for each refinery 12 or for each type of rotating machine 14. The tentative learning data group 52 includes normal data D1, abnormal data D2, or simulated abnormal data D3.

[0051] The main learning data group 54 is a collection of learning data used for main learning. Here, "main learning" refers to additional or re-learning performed after tentative learning, and refers to individual learning performed for each rotating machine 14. Like the tentative learning data group 52, the main learning data group 54 includes normal data D1, abnormal data D2, or simulated abnormal data D3. Here, the simulated abnormal data D3 is data that simulates the abnormal data D2, and is used in the learning process or verification process for the anomaly detection model M.

[0052] The normal data D1, the abnormal data D2, and the simulated abnormal data D3 are all time-series data of physical quantities correlated with the vibration of the rotating machine 14. The sampling periods of these data are all shaped to be equal. Furthermore, these time-series data are classified according to whether they have been processed or not, and whether or not an abnormality has occurred.

[0053] "Whether or not the data is processed" means whether or not the raw data measured by the sensors 20-22 is used as is without substantial modification of its content. Specifically, the normal data D1 and the abnormal data D2 both correspond to unprocessed time-series data. In contrast, the simulated abnormal data D3 corresponds to processed time-series data.

[0054] "Presence or absence of an abnormality occurrence interval" means whether or not there is an interval indicating that an abnormality has occurred in the state of the rotating machine 14 due to an event such as a failure (i.e., an abnormality occurrence interval). Specifically, the normal data D1 corresponds to time-series data in which all intervals of the time-series data are "normal intervals" and do not include "abnormality occurrence intervals." A "normal interval" refers to an interval in which it is considered that no abnormality has occurred in the rotating machine 14. A "normal interval" refers to an interval in which a person (e.g., an operator) has determined that no abnormality has occurred in the rotating machine 14 (i.e., it is normal) based on the time change of physical quantities included in the data. In contrast, the abnormality data D2 and the simulated abnormality data D3 correspond to time-series data that includes at least one "abnormality occurrence interval" among all intervals of the time-series data.

[0055] The abnormality occurrence section includes a transient section. That is, the abnormality occurrence section is either [1] a transient section only, [2] a combination of a transient section and a normal section, [3] a combination of a transient section and an abnormal section, or [4] a combination of a normal section, a transient section, and an abnormal section. Here, the "transient section" refers to the section from when the state of the rotating machine 14 transitions from normal to abnormal.

[0056] The transient section includes a "time point at which an abnormality occurs," which is a time point at which a person (e.g., a user) can determine that an abnormality has occurred in the rotating machine 14 from the time change of a physical quantity included in the data, as well as a period before the time point at which the abnormality occurs and a period after the time point at which the abnormality occurs. The "time point at which the abnormality occurs" is set in advance by a user for each time series data, for example, before performing a learning process or a verification process for the anomaly detection model M. The "period before the time point at which the abnormality occurs" in the transient section is a period during which an abnormality in the rotating machine 14 is suspected, and corresponds to a period including the time point at which the actual abnormality occurred, which the person who determined the time point at which the abnormality occurred was unable to fully determine. In other words, the "period before the time point at which the abnormality occurred" in the transient section is a period including an incubation period until an abnormality that has occurred in the rotating machine 14 becomes apparent enough for a person to recognize the occurrence of the abnormality. The "period after the time point at which the abnormality occurs" in the transient section is a period during which an abnormality has already occurred in the rotating machine 14.

[0057] The tentative learning parameter group 56 is a collection of learning parameters identified through tentative learning. The tentative learning parameter group 56 is determined, for example, for each refinery 12 registered in the state determination device 18 or for each type of rotating machine 14.

[0058] The main learning parameter group 58 is a collection of learning parameters identified through main learning. The main learning parameter group 58 is determined for each rotating machine 14 registered in the state determination device 18, for example.

[0059] Examples of learning parameters include [1] "model parameters" (or hyperparameters) for specifying the model structure of the anomaly detection model M, or [2] "variable parameters" whose optimal values ​​change depending on the population of training data. Examples of variable parameters include weight coefficients between calculation units, thresholds for activation functions, and principal component coefficients.

[0060] The determination target data D4 is time-series data used to determine the state of the rotating machine 14. Here, the determination target data D4 is time-series data obtained through measurements by the sensors 20 to 22 after the main learning parameter group 58 is determined through main learning.

[0061] The judgment result information 60 includes [1] "time information" such as the date and time, [2] "identification information" for identifying the rotating machine 14, [3] "meta information" such as the type of rotating machine 14 and the name of the refinery 12, or [4] "judgment results" such as a flag value indicating whether the condition is normal or abnormal, and the magnitude of the judgment error.

[0062] <Explanation of anomaly detection model M> Next, the anomaly detection model M included in the data determination unit 48 in Fig. 2 will be described with reference to Fig. 3 and Fig. 4. A time series of physical quantities (hereinafter referred to as time series feature quantities) is used as an input value of the anomaly detection model M. One example of a method for generating time series feature quantities is "partial time series generation."

[0063] FIG. 3 is a diagram schematically illustrating an example of a method for generating time series feature quantities. The horizontal axis of the graph represents time, and the vertical axis of the graph represents physical quantities. Here, it is assumed that the reference time point is set to T0, the sampling period is set to Δt, and the window width is set to "5." In this case, a plurality of partial time series data are generated from one piece of time series data by sequentially sliding the extraction interval starting from the reference time point t=T0 (the 0th sample point). These plurality of partial time series data are generated using, for example, normal data D1, abnormal data D2, simulated abnormal data D3, or judgment target data D4.

[0064] Specifically, the first partial time series data is composed of the 0th to 4th sample value sets. The second partial time series data is composed of the 1st to 5th sample value sets. The third partial time series data is composed of the 2nd to 6th sample value sets. The fourth partial time series data is composed of the 3rd to 7th sample value sets. The fifth partial time series data is composed of the 4th to 8th sample value sets. In this way, through preprocessing of the time series data (partial time series generation), time series feature quantities, which are input values ​​for the anomaly detection model M, are generated.

[0065] Fig. 4 is a functional block diagram relating to the anomaly detection model M shown in Fig. 2. This anomaly detection model M is configured to include an unsupervised learner 70, an error calculation unit 72, an error combination unit 74, and a threshold processing unit 76.

[0066] The unsupervised learning device 70 performs a "reconstruction operation" to reconstruct time-series features related to physical quantities through dimensionality reduction and dimensionality restoration. Examples of the reconstruction operation include principal component analysis and autoencoder. The unsupervised learning device 70 inputs original time-series features 80 and outputs reconstructed time-series features 82.

[0067] The original time series feature quantity 80 is the partial time series data described in Fig. 3 or a feature quantity generated from the partial time series data. The time series feature quantity 80 is composed of a feature quantity (V) related to vibration velocity, a feature quantity (A) related to vibration acceleration, and a feature quantity (T) related to temperature. Note that, because the definitions of the three types of physical quantities are different, standardization or normalization is performed to match the scale of each data.

[0068] The reconstructed time-series feature 82 is a feature obtained by reconstructing the original time-series feature 80 through the unsupervised learning device 70. The time-series feature 82 is composed of a feature (V') related to vibration velocity, a feature (A') related to vibration acceleration, and a feature (T') related to temperature.

[0069] The error calculation unit 72 performs a "calculation operation" to find the difference between each component of the time-series feature quantities 80 and 82 and calculate a reconstruction error for each physical quantity. This reconstruction error may be either the mean absolute error (MAE) or the root mean square error (RMSE), and is standardized or normalized for each physical quantity. Hereinafter, the standardized or normalized reconstruction errors for the vibration velocity, vibration acceleration, and temperature will be referred to as the first error Er1, the second error Er2, and the third error Er3, respectively.

[0070] The error combining unit 74 performs a "combining operation" to combine the multiple reconstruction errors calculated by the error calculation unit 72. The combining operation includes an operation to calculate statistics such as an average value, a weighted average value, a maximum value, and a minimum value. When calculating the weighted average value, weighting coefficients W1, W2, and W3 corresponding to the first error Er1, the second error Er2, and the third error Er3 are set in advance. The magnitude relationship between W1 to W3 can be set arbitrarily, but satisfies, for example, W1=W2>W3>0. Hereinafter, the combined reconstruction error will be referred to as an error index Eidx.

[0071] The threshold processing unit 76 performs a "threshold calculation" to output a flag value indicating whether the result is normal or abnormal, depending on the magnitude relationship between the error index Eidx combined by the error combining unit 74 and the threshold value Th corresponding to the error index Eidx. For example, when the error index Eidx exceeds the threshold value Th, the threshold processing unit 76 outputs a flag value indicating an abnormal determination result. For example, when the error index Eidx is equal to or smaller than the threshold value Th, the threshold processing unit 76 outputs a flag value indicating a normal determination result. This flag value is defined, for example, so that 0 indicates "normal" and 1 indicates "abnormal."

[0072] [Operation of the state determination system 10] The state determination system 10 in this embodiment is configured as described above. Next, an outline of the operation of the state determination system 10 will be described with reference to FIGS.

[0073] <1. Overall operation> Fig. 5 is a sequence diagram relating to the operation of the condition determination system 10 shown in Fig. 1. Each step in this sequence diagram is executed by the process monitoring device 16 and the condition determination device 18 working together.

[0074] In step SP10, the state determination device 18 collects learning data to be used for tentative learning from the process monitoring devices 16 installed in various refineries 12. As a result, a tentative learning data group 52 is formed.

[0075] In step SP12, the data acquisition unit 40 of the state determination device 18 reads and acquires the tentative learning data group 52 from the storage unit 34 when the timing for tentative learning arrives.

[0076] In step SP14, if there are insufficient positive examples (i.e., abnormal cases) in the abnormality detection model M, the data generation unit 42 of the state determination device 18 generates simulated abnormal data D3 from the normal data D1 and abnormal data D2 as necessary.

[0077] In step SP16, the learning processing unit 44 of the state determination device 18, in cooperation with the model evaluation unit 46, performs tentative learning for each refinery 12 or each type of rotating machine 14 using the tentative learning data group 52. Through this learning process, a tentative learning parameter group 56 is determined.

[0078] In step SP18, the state determination device 18 stores in the storage unit 34 the tentative learning parameter group 56 determined through the tentative learning in step SP16.

[0079] In step SP20, after the rotating machine 14 to be monitored is newly installed, the process monitoring device 16 periodically or irregularly transmits time-series data of physical quantities contained in the measurement signals from the sensors 20 to 22 to the state determination device 18. As a result, a main learning data group 54 to be used for main learning is accumulated.

[0080] In step SP22, the data acquisition unit 40 of the state determination device 18 reads and acquires the main learning data group 54 from the storage unit 34 when the time for main learning arrives.

[0081] In step SP24, if there are no or insufficient positive examples (i.e., abnormal cases) in the abnormality detection model M, the data generation unit 42 of the state determination device 18 generates simulated abnormal data D3 using normal data D1 belonging to the main learning data group 54 and abnormal data D2 belonging to the provisional learning data group 52.

[0082] In step SP26, the learning processing unit 44 of the state determination device 18 performs additional or re-learning processing (i.e., main learning) on ​​the anomaly detection model M using the main learning data group 54 and the temporary learning parameter group 56. Through this learning processing, a main learning parameter group 58 is determined.

[0083] In step SP28, the state determination device 18 stores in the storage unit 34 the main learning parameter group 58 determined through the learning process in step SP26.

[0084] In step SP30, the process monitoring device 16 periodically or irregularly transmits time-series data of physical quantities contained in the measurement signals from the sensors 20 to 22 to the state determination device 18. As a result, the state determination device 18 acquires determination target data D4.

[0085] In step SP32, the data determination unit 48 of the state determination device 18 determines the state of the rotating machine 14 by inputting the determination target data D4 acquired in step S30 into the abnormality detection model M specified by the main learning parameter group 58. Through this determination process, determination result information 60 is generated.

[0086] In step SP34, the output instruction section 50 of the state determination device 18 transmits data including the determination result information 60 obtained in step SP32. As a result, the determination result information 60 is output via the process monitoring device 16.

[0087] The state determination device 18 may re-learn the anomaly detection model M when the timing for re-learning the anomaly detection model M arrives. When the state determination device 18 re-learns the anomaly detection model M, the state determination device 18 may perform the re-learning in the same manner as the processing shown in steps SP22 to SP28 above. Examples of the timing for re-learning include: [1] when the state determination device 18 receives an instruction to perform re-learning from a terminal (not shown) owned by the user of the state determination system 10 or the operator of the rotating machine 14; [2] when the number of erroneous determinations by the anomaly detection model M exceeds a predetermined number; or [3] when a predetermined time has elapsed since the processing of step SP28 (including saving the learning parameter group 58 resulting from re-learning in the memory unit 34) was last executed.

[0088] 6 is a diagram showing an example of the determination result of abnormal data D2 by anomaly detection model M. The horizontal axis of the graph indicates time (unit: days), and the vertical axis of the graph indicates [1] vibration acceleration (shown by a solid line) or [2] reconstruction error (shown by a dashed line). In the example of FIG. 6, only seven sections B1 to B7 of the entire time series are partially shown.

[0089] Of the sections B1 to B7, sections B1 to B3 and B5 to B7 all correspond to "normal sections." Within the normal sections, the vibration acceleration and the reconstruction error each tend to remain at relatively small values, although there are some fluctuations. Note that in the example shown in FIG. 6, even within the normal section, the reconstruction error exceeds the threshold, resulting in several "misjudgments" (or false positive judgments) in which the state of the rotating machine 14 is judged to be abnormal.

[0090] Of the sections B1 to B7, section B4 corresponds to the "abnormality occurrence section." Within the abnormality occurrence section, the vibration acceleration and reconstruction error each tend to gradually increase in the transient section until they transition from normal to abnormal, remaining at relatively large values. Furthermore, within the abnormality section, the reconstruction error continues to exceed the threshold. Note that the behavior within the abnormality occurrence section has various time series patterns, including not only the above-mentioned rising pattern but also a falling pattern.

[0091] Incidentally, when the rotating machine 14 is operating, vibrations emitted by peripheral devices may be measured by the sensors 20 and 21, which may result in a false determination that the rotating machine 14 is abnormal even though the state of the rotating machine 14 itself is normal. In this case, the false determination occurs because the state determination device 18 does not determine the state of the rotating machine 14 based solely on the vibrations emitted by the rotating machine 14 itself.

[0092] Here, the state of the rotating machine 14 may be reflected not only by time changes in vibration-related physical quantities but also by time changes in other physical quantities. The state of the rotating machine 14 may also be reflected, for example, by time changes in temperature. For example, if the vent of the bearing housing of the rotating machine 14 becomes blocked, the surface temperature of the bearing housing increases due to frictional heat generated by the bearings housed in the bearing housing. Therefore, by including temperature in addition to vibration velocity and vibration acceleration in the time-series feature quantity 80 shown in FIG. 4, the state of the rotating machine 14 can be determined from multiple perspectives. This further improves the accuracy of determining the state of the rotating machine 14 in the above-mentioned cases. Note that when the state determination device 18 determines the state of the rotating machine 14 based on temperature, there is a possibility that a spike in the temperature measured by the sensor 22 due to rainfall, for example, may lead to an erroneous determination. Therefore, the user may appropriately adjust the influence of temperature on the determination results of the anomaly detection model M depending on the surrounding environment of the rotating machine 14 (monitored object).

[0093] For example, by relatively reducing the contribution of temperature in the anomaly detection model M, it is possible to suppress erroneous determinations caused by spike-like temperature fluctuations. One example of a method for reducing the contribution of temperature compared to the vibration velocity and vibration acceleration is to set weighting coefficients so that W1 > W3 and W2 > W3 (see FIG. 4 ). In the above example, the state of the rotating machine 14 is reflected in the surface temperature. However, when the state determination device 18 determines the state of the monitored object based on temperature, the temperature used for the determination is not limited to the surface temperature of the rotating machine 14. For example, the internal temperature of the rotating machine 14, the ambient temperature of the rotating machine 14, or a combination of these temperatures may be used. Furthermore, the state determination device 18 may determine the state of the rotating machine 14 based on a time change in a physical quantity other than temperature.

[0094] <2. Generation of simulated abnormal data D3> Next, a method for generating the simulated abnormality data D3 (steps SP14 and SP24 in FIG. 5) will be described in detail with reference to FIGS.

[0095] Fig. 7 is a functional block diagram relating to the data generation unit 42 shown in Fig. 2. Specifically, the data generation unit 42 is configured to include an abnormal section extraction unit 90, a pattern calculation unit 92, a processing section extraction unit 94, and a data multiplication unit 96.

[0096] The abnormality section extraction unit 90 extracts, from the entire time series of the abnormality data D2, a section indicating that an abnormality has occurred in the state of the rotating machine 14 (i.e., an abnormality occurrence section). The abnormality section extraction unit 90 may, for example, analyze log data that records the operating state of the rotating machine 14 and refer to the time point when the abnormality occurred in the rotating machine 14 to extract the corresponding abnormality occurrence section. Note that the abnormality occurrence section includes a transition section in which the state of the rotating machine 14 transitions from normal to abnormal. The reason for this is that the abnormality occurrence section expresses not only the "state in which the rotating machine 14 has become abnormal" but also "a sign that the state of the rotating machine 14 will become abnormal."

[0097] The pattern calculation unit 92 calculates a modulation pattern PT to be used for generating the simulated abnormality data D3, using sample values ​​belonging to the abnormality occurrence section extracted by the abnormality section extraction unit 90. This modulation pattern PT indicates a temporal change in the relative intensity of vibrations that actually occurred as an "abnormality" in the actual machine. The modulation pattern PT can be calculated, for example, by dividing each sample value of a physical quantity belonging to the abnormality occurrence section by a representative value of a physical quantity belonging to a normal section of the entire time series, excluding the abnormality occurrence section. This representative value can be, for example, a statistic (such as the mean, median, or mode) of sample values ​​in a data population obtained when the rotating machine 14 is in a low-deterioration state (more specifically, within a predetermined period from the start of operation of the rotating machine 14).

[0098] The processing section extraction unit 94 extracts a section to be processed (i.e., a processing section) from the entire section indicated by the time series data to be processed (here, normal data D1). This processing section is set so as to have the same length as the abnormality occurrence section. In other words, the number of samples of the physical quantity belonging to the processing section matches the number of data of the modulation pattern PT.

[0099] This processing section may be any section belonging to the normal section, but may be set to a section that is easy to reproduce the behavior in the abnormality section, for example. Specifically, the processing section extraction unit 94 may take a moving average of the time-series data and set the processing section so that it includes a position where the level is relatively high. This makes it possible to prevent this specific section from being extracted as the processing section, even if the vibration temporarily decreases due to the operation of the inverter that drives the rotating machine 14.

[0100] The data multiplication unit 96 multiplies the normal data D1, whose processing interval has been extracted by the processing interval extraction unit 94, by the modulation pattern PT calculated by the pattern calculation unit 92. Specifically, the data multiplication unit 96 sequentially multiplies the i-th (i = 1 to N) sample value by the corresponding i-th relative value, thereby generating simulated abnormal data D3.

[0101] 8 is a detailed flowchart of the method for generating the simulated abnormality data D3 (steps SP14 and SP24 in FIG. 6). Here, an example will be described in which the physical quantity correlated with vibration is "vibration acceleration."

[0102] In step SP40, the data generating unit 42 acquires normal data D1 to be processed and abnormal data D2 to be used in processing the normal data D1.

[0103] In step SP42, the data generation unit 42 (more specifically, the abnormal section extraction unit 90) analyzes the abnormal data D2 acquired in step SP40 and extracts an abnormality occurrence section from all sections of the time series. In the example of the abnormal data D2 shown in Figure 6, section B4 is extracted.

[0104] In step SP44, the data generating section 42 (more specifically, the pattern calculating section 92) calculates a modulation pattern PT indicating the change over time in the relative strength of the vibration acceleration from each sample value in the abnormality occurrence section extracted in step SP42.

[0105] 9 is a diagram showing an example of the calculation results of the modulation pattern PT. The horizontal axis of the graph indicates time (unit: days), and the vertical axis of the graph indicates [1] vibration acceleration (unit: m / s2) or [2] relative value (unit: dimensionless). The vibration acceleration in the abnormal data D2 is maintained at a relatively low and approximately constant value in a "normal" state, but gradually increases over time in a "transient" state, and is maintained at a relatively high and approximately constant value in an "abnormal" state. As can be seen from FIG. 9, the modulation pattern PT has a waveform similar to the behavior of the vibration acceleration described above.

[0106] In step SP46 of FIG. 8, the data generating section 42 (more specifically, the processing section extracting section 94) analyzes the normal data D1 acquired in step SP40, and extracts processing sections from all sections of the time series.

[0107] In step SP48, the data generating section 42 (more specifically, the data multiplying section 96) multiplies each sample value of the vibration acceleration belonging to the processing section extracted in step SP46 by the modulation pattern PT calculated in step SP44.

[0108] 10A and 10B are diagrams showing an example of the processing result of normal data D1. More specifically, FIG. 10A shows the normal data D1 before processing, and FIG. 10B shows the normal data D1 after processing (i.e., simulated abnormal data D3). As can be seen from FIG. 10, the simulated abnormal data D3 exhibits a behavior of gradually increasing in the transition section within the processing section until it transitions from normal to abnormal.

[0109] 8, the data generation unit 42 stores the simulated abnormality data D3 generated through the multiplication in step SP46 in the storage unit 34. Thereafter, the simulated abnormality data D3 becomes available for use in the learning process or verification process of the anomaly detection model M.

[0110] In this way, the data generator 42 generates the simulated abnormal data D3 from the normal data D1 and the abnormal data D2 according to the flowchart of FIG. 8 (steps SP14 and S24 in FIG. 5).

[0111] <3. Evaluation of anomaly detection model M> Next, the evaluation method for the anomaly detection model M, which is executed in step SP16 of FIG. 5, will be described in detail with reference to FIGS.

[0112] Prior to evaluation of the anomaly detection model M, the entire interval in the normal data D1 or the abnormal data D2 is divided in advance into a "learning period" used in the learning process and a "verification period" used in the verification process. For this verification period (or judgment period), for example, a normal interval closest to the current time is selected. Then, a portion of this verification period is set as a processing interval, and simulated abnormal data D3 having a simulated abnormality occurrence interval is generated.

[0113] Fig. 11 is a functional block diagram relating to the model evaluation unit 46 shown in Fig. 2. Specifically, the model evaluation unit 46 is configured to include a data analysis unit 100, a parameter designation unit 102, an index calculation unit 104, and an index evaluation unit 106, in addition to the anomaly detection model M that has undergone learning processing.

[0114] The data analysis unit 100 performs an analysis process on time-series data used to evaluate the anomaly detection model M, here, the normal data D1, the abnormal data D2, or the simulated abnormal data D3, to obtain the time transition of the actual results regarding whether the state of the rotating machine 14 is normal or abnormal. The data analysis unit 100 may, for example, analyze log data that records the operating state of the rotating machine 14, to obtain the time transition of the state of the rotating machine 14.

[0115] The parameter specification unit 102 specifies one or more model parameters for identifying the model structure of the anomaly detection model M, and supplies the model parameters to the anomaly detection model M and the index evaluation unit 106. By repeatedly specifying and evaluating the model parameters, a so-called "grid search" is performed.

[0116] Examples of model parameters include: [1] the number and length of data of the time series feature 80 (Figure 4); [2] the number of principal components in the unsupervised learning device 70 (principal component analysis); [3] hyperparameters related to the network structure of the unsupervised learning device 70 (autoencoder); [4] the type of reconstruction error in the error calculation unit 72; [5] the weighting coefficient in the error synthesis unit 74; or [6] the magnitude of the threshold in the threshold processing unit 76.

[0117] The anomaly detection model M is a mathematical model that has been trained in accordance with model parameters specified by the parameter specification unit 102. The trained anomaly detection model M sequentially inputs time-series feature quantities 80 generated from the abnormal data D2 or the simulated abnormal data D3, and outputs the time transition of the determination result as to whether the state of the rotating machine 14 is normal or abnormal.

[0118] The index calculation unit 104 calculates an evaluation index related to the detection performance of the anomaly detection model M, using the time transition of the actual results obtained through the data analysis unit 100 and the time transition of the judgment results obtained through the trained anomaly detection model M. The index calculation unit 104 is configured to include, for example, an erroneous judgment count calculation unit 108 and a grace period calculation unit 110.

[0119] The erroneous determination count calculation unit 108 calculates the number of times an abnormality is determined to exist within a determination period that does not include an abnormality occurrence section (hereinafter referred to as the "number of erroneous determinations"), with respect to the time transition of the determination results. The smaller the number of erroneous determinations, the higher the detection performance of the anomaly detection model M is evaluated to be. The number of erroneous determinations may be an actual count value or a count value compiled in accordance with actual operation. "Compiled in accordance with actual operation" means, for example, when notification to the process monitoring device 16 is made daily, counting two or more erroneous determinations within one day as one erroneous determination. The erroneous determination count calculation unit 108 may, for example, count the number of times an abnormality is determined to exist for normal data D1 as the number of erroneous determinations. Furthermore, the erroneous determination count calculation unit 108 may, for example, count the number of times an abnormality is determined to exist for abnormal data D2, the determination period of which is set so as not to include an abnormality occurrence section, as the number of erroneous determinations. Furthermore, the erroneous determination count calculation unit 108 may count, as the number of erroneous determinations, the number of times that simulated abnormality data D3, for which a determination period is set so as not to include the abnormality occurrence section, is determined to be abnormal.

[0120] The grace time calculation unit 110 calculates the time (hereinafter, "grace time") from the point at which an abnormality is first determined to exist in the rotating machine 14 during the determination period including the abnormality occurrence section to the point at which the abnormality occurs, with respect to the time transition of the determination result. The larger the grace time, the higher the detection performance of the abnormality detection model M is evaluated to be. When there are two or more abnormality occurrence points within the determination period, the grace time calculation unit 110 calculates the grace time for the abnormality occurrence point of interest based on the point at which the first abnormality determination was made that is after the immediately preceding abnormality occurrence point and before the abnormality occurrence point of interest. In this case, the grace time calculation unit 110 may calculate the grace time for the first abnormality occurrence point within the determination period based on the point at which the first abnormality determination was made before the abnormality occurrence point. The grace time calculation unit 110 may calculate the grace time for the abnormality data D2 or the simulated abnormality data D3.

[0121] The index evaluation unit 106 evaluates the number of misjudgments and the grace period calculated for each model parameter. The index evaluation unit 106 selects and determines one model parameter from among multiple model parameter candidates in accordance with predetermined selection rules. This selection rule may be determined by [Rule 1] only the number of misjudgments, [Rule 2] only the grace period, or [Rule 3] a combination of the number of misjudgments and the grace period. An example of rule 1 is that the number of misjudgments is minimized. An example of rule 2 is that the grace period is maximized. An example of rule 3 is that [1] the number of misjudgments is below a threshold and the grace period is longest, or [2] the grace period is longer than the threshold and the number of misjudgments is smallest.

[0122] 12 is a detailed flowchart of a method for evaluating the anomaly detection model M associated with the tentative learning (step SP16 in FIG. 5). Here, the determination of model parameters using a grid search will be described as an example.

[0123] In step SP60, the model evaluation unit 46 acquires the provisional training data group 52 (here, the abnormal data D2 or the simulated abnormal data D3) to be used for evaluation.

[0124] In step SP62, the data analysis section 100 analyzes the abnormal data D2 or the simulated abnormal data D3 acquired in step SP60, thereby obtaining the time transition of the actual results.

[0125] In step SP64, the parameter specification unit 102 specifies a candidate value set for model parameters that have not yet been specified. This candidate value set is supplied to the anomaly detection model M and the index evaluation unit 106, respectively.

[0126] In step SP66, the anomaly detection model M executes an anomaly detection simulation using the anomaly data D2 or simulated anomaly data D3 acquired in step SP60, thereby obtaining the time transition of the determination result.

[0127] In step SP68, the index calculation unit 104 calculates an evaluation index regarding the detection performance of the anomaly detection model M by using the time transition of the actual result obtained in step SP64 and the time transition of the determination result obtained in step SP66. Specifically, [1] calculation of the number of misjudgments by the misjudgment number calculation unit 108 and [2] calculation of the lead time by the lead time calculation unit 110 are respectively performed.

[0128] FIG. 13 is a diagram showing an example of a method for calculating the number of misjudgments. The horizontal axis of the graph shows time respectively. As shown in the upper part of FIG. 13, in the entire time series range, the time range of T1≦t≦T2 is set as the "learning period", and the time range of T2≦t≦T3 is set as the "verification period". The time transition of the actual result is shown in the middle part of FIG. 13, and the time transition of the determination result is shown in the lower part of FIG. 13. Here, the number of misjudgments is calculated within the verification period consisting only of the "normal section". That is, it is assumed that no anomaly occurs at all in the time range of T2≦t≦T3 with respect to the time transition of the actual result. According to the time transition of the determination result, misjudgments are made at t = Af1 to Af5 during the period of T2≦t≦T3. In this case, the number of misjudgments is obtained as "5 times".

[0129] FIG. 14 is a diagram showing an example of a method for calculating the lead time. The horizontal axis of the graph shows time respectively. As shown in the upper part of FIG. 14, in the entire time series range, the time range of T1≦t≦T2 is set as the "learning period", and the time range of T2≦t≦T3 is set as the "verification period". The time transition of the actual result is shown in the middle part of FIG. 14, and the time transition of the determination result is shown in the lower part of FIG. 14. Here, the number of misjudgments is calculated within the verification period including the pseudo-generated "anomaly occurrence section". That is, it is assumed that a person can first recognize the anomaly that occurred in the rotating machine 14 at the anomaly occurrence time point t = Td (T2 < Td < T3) with respect to the time transition of the actual result. According to the time transition of the determination result, anomaly determinations are made at t = Af1 to Af5 until the last determination is made at the time point t = Td (time point t = At) (until the anomaly occurrence time point and the determination time point match). In this case, the lead time is obtained as (At - Af1).

[0130] 12, the parameter specification unit 102 checks whether or not all candidate value sets have been specified. If not all candidate value sets have been specified (step SP70: NO), the parameter specification unit 102 returns to step SP64 and specifies another candidate value set.

[0131] Thereafter, the model evaluation unit 46 sequentially repeats steps SP64 to SP68 to calculate the number of erroneous determinations and the grace period for each candidate value set. Returning to step SP70, if all candidate value sets have been specified (step SP70: YES), the model evaluation unit 46 proceeds to the next step SP72.

[0132] In step SP72, the index evaluation unit 106 performs grouping by refinery 12 or by type of rotating machine 14, and calculates statistics of the evaluation index for each group.

[0133] In step SP74, the index evaluation unit 106 determines model parameters (i.e., part of the provisional learning parameter group 56) for each refinery 12 or each type of rotating machine 14 using the statistics of the evaluation index calculated in step SP72.

[0134] FIG. 15 is a diagram schematically illustrating an example of a method for determining model parameters. The horizontal axis of the graph indicates the number of erroneous determinations (unit: times), and the vertical axis of the graph indicates the grace period (unit: days). In the example of FIG. 15, evaluation levels corresponding to six candidate value sets are plotted as P1 to P6, respectively. Group G1 is made up of all evaluation levels P1 to P6 belonging to the population. Group G2 is made up of evaluation levels P1 to P3 belonging to group G1 whose number of erroneous determinations is below threshold Nth. Of the evaluation levels P1 to P3 belonging to group G2, the evaluation level with the longest grace period is selected. As a result, evaluation level P1 is determined as the determined value set of the model parameters.

[0135] In this way, the model evaluation unit 46 determines the provisional learning parameter set 56 according to the flowchart of FIG. 12 (step SP16 in FIG. 5).

[0136] [Summary of the embodiment] As described above, the state determination device 18 generates simulated abnormality data D3 by multiplying each sample value of the physical quantity belonging to the processing section in the data to be processed by the modulation pattern PT indicating the time change of the relative strength of vibration in the abnormality occurrence section. This makes it possible to generate simulated time series data that reproduces the time change of the actual physical quantity in the abnormality occurrence section (here, appropriate vibration behavior with an upward slope), and even when time series data that does not include the abnormality occurrence section is obtained, it is possible to construct an abnormality detection model M that has high accuracy in determining whether something is normal or abnormal through learning processing.

[0137] Alternatively, the modulation pattern PT may be calculated by dividing each sample value of the physical quantity belonging to the abnormality occurrence section by a representative value of the physical quantity belonging to the normal section of the time series, excluding the abnormality occurrence section, thereby improving the smoothness at the connection point between the normal section and the abnormality section.

[0138] The abnormality occurrence section may also include a transition section in which the state of the rotating machine 14 transitions from normal to abnormal. This makes it possible to reproduce, through the simulated abnormality data D3, a sign that the state of the rotating machine 14 will become abnormal.

[0139] Furthermore, the processing section may be a section in which the physical quantity is relatively large compared to other sections among all sections of the time series. By making the value of the multiplicand relatively large, the reproducibility of the time change of the actual physical quantity in the abnormality occurrence section is further improved.

[0140] Moreover, the state determination device 18 may further include a model evaluation unit 46 that calculates an evaluation index related to the detection performance of the anomaly detection model M and evaluates the anomaly detection model M based on the evaluation index.

[0141] The first evaluation index may include the number of times an abnormality is determined to exist within a determination period that does not include an abnormality occurrence section (i.e., the number of incorrect determinations) with respect to the time transition of the determination result. This allows quantification of detection performance from the viewpoint of normal sample accuracy (or specificity).

[0142] The second evaluation index may include, with respect to the time transition of the judgment result, the time from the point at which the judgment of an abnormality is first made to the point at which the abnormality occurs within the judgment period including the abnormality occurrence section (i.e., grace period). This allows quantification of the detection performance from the viewpoint of the accuracy (or sensitivity) of the abnormality sample.

[0143] Alternatively, the model evaluation unit 46 may calculate the number of erroneous determinations and the grace period corresponding to each of the multiple model parameter candidates, and select one model parameter candidate from the multiple model parameter candidates according to the combination of the number of erroneous determinations and the grace period. This allows the selection of model parameters that take into account both the accuracy of normal samples and the accuracy of abnormal samples.

[0144] Alternatively, the model evaluation unit 46 may select one model parameter candidate that reduces the number of erroneous judgments below a threshold and provides the longest grace period. This maximizes the time allowance from the point at which a sign of an abnormality is detected while still ensuring a predetermined normal sample accuracy.

[0145] Furthermore, when the monitored object is a rotating machine 14 installed in a facility (here, a refinery 12), the learning processing unit 44 may perform additional or re-learning processing on an anomaly detection model M provisionally determined for each refinery 12 or each type of rotating machine 14, thereby finally determining an anomaly detection model M to be applied to each individual rotating machine 14. In this way, an anomaly detection model M customized for each rotating machine 14 can be constructed.

[0146] Furthermore, the physical quantities may include the vibration velocity, vibration acceleration, and temperature of the rotating machine 14. In this case, the error index is a linear sum of a first error Er1 of the vibration velocity, a second error Er2 of the vibration acceleration, and a third error Er3 of the temperature, and the weighting coefficient W3 of the third error Er3 may be set to be smaller than the weighting coefficients W1 and W2 of the first error Er1 and the second error Er2. This allows for more accurate determination including temperature, and also makes it possible to suppress erroneous determinations caused by spike-like temperature fluctuations.

[0147] The time-series feature 80 may also include multiple partial time-series data extracted from the time-series data of physical quantities by sequentially sliding an extraction interval starting from a reference time point, thereby generating feature values ​​suitable for, for example, principal component analysis.

[0148] [Variations] The present disclosure is not limited to the above-described embodiments, and can be freely modified without departing from the spirit of the present disclosure. Alternatively, the respective configurations may be arbitrarily combined within the scope of no technical contradiction. Alternatively, the execution or execution order of each step constituting a flowchart or sequence diagram may be changed within the scope of no technical contradiction.

[0149] In the above embodiment, an example has been described in which the state of the rotating machine 14 is determined, but the object to be monitored is not limited to the rotating machine 14 as long as it is an object that vibrates. Furthermore, the location where the rotating machine 14 is installed is not limited to the refinery 12, and it may be various indoor or outdoor facilities such as a chemical plant, a steel mill, or a power plant.

[0150] In the above embodiment, the simulated abnormal data D3 is generated from normal data D1 without an abnormality occurrence interval, but the simulated abnormal data D3 may be generated from abnormal data D2 with an abnormality occurrence interval. In this case, the data generator 42 extracts a processing interval from the normal interval included in the time series of the abnormal data D2, and applies a modulation pattern PT calculated from other abnormal data D2 to the processing interval.

[0151] In the above embodiment, the case where the sensors 20, 21, and 22 are attached to the rotating machine 14 has been described as an example, but this is not limiting. As an example, the sensor 22 may be omitted. In this case, each process performed by the control unit 32 on the time-series temperature data is omitted.

[0152] In the above embodiment, the process monitoring device 16 and the status determination device 18 do not have to be able to communicate with each other via the network NT. For example, the process monitoring device 16 may be able to communicate unidirectionally from the process monitoring device 16 to the status determination device 18 via the network NT. In this case, the status determination device 18 may transmit (output) the determination result via the network NT to a terminal (not shown) owned by, for example, a user of the status determination system 10 or an operator of the rotating machine 14.

[0153] The information, physical quantities, feature quantities, sample values, indicators, parameters, etc. described in the present disclosure may be expressed using absolute values, may be expressed using relative values ​​from a predetermined value, or may be expressed using corresponding other information.

[0154] The term "determining" in this disclosure may encompass a wide variety of actions. "Determining" may include, for example, determining, calculating, computing, processing, deriving, investigating, looking up (e.g., looking up in a table, database, or other data structure), ascertaining, and so forth. "Determining" may also include, for example, resolving, selecting, choosing, establishing, comparing, and so forth. That is, "determining" may include ascertaining, ascertaining, some action.

[0155] In this disclosure, when expressions such as "obtaining / setting / using / based on / using information as input" (including similar expressions) are used, unless otherwise specified, this includes cases where the information itself is used, or where information that has been processed in some way (e.g., information that has been noise-added, normalized, features extracted from the information, intermediate representations of the information, etc.) is used. Furthermore, when a statement is made that a result is obtained "by obtaining / setting / using / based on / using information as input" (including similar expressions), this includes cases where the result is obtained based solely on the information, or cases where the result is influenced by other information, factors, conditions, and / or states other than the information in question, unless otherwise specified. Furthermore, when a statement is made that "outputting information" (including similar expressions), this includes cases where the information itself is used as output, or where information that has been processed in some way (e.g., information that has been noise-added, normalized, features extracted from the information, intermediate representations of various information, etc.) is used as output, unless otherwise specified. [Explanation of symbols]

[0156] 10. Condition determination system, 12. Refinery (facility), 14. Rotating machine (object to be monitored), 16. Process monitoring device, 18. Condition determination device, 20-22. Sensor, 40. Data acquisition unit, 42. Data generation unit, 44. Learning processing unit, 46. Model evaluation unit, 48. Data determination unit, 50. Output instruction unit, 52. Temporary learning data group, 54. Actual learning data group, D1. Normal data, D2. Abnormal data, D3. Simulated abnormal data, D4. Data to be determined, M. Anomaly detection model

Claims

1. a data acquisition unit that acquires abnormality data, which is time-series data of physical quantities correlated with vibrations of a monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the state of the monitored object within the entire section of the time series, and processing target data, which is time-series data of the physical quantities and is an object to be processed; a pattern calculation unit that calculates a modulation pattern indicating a time change in the relative intensity of the vibration in the abnormality occurrence section using the abnormality data acquired by the data acquisition unit; and a data generating unit that extracts a processing section equal to the length of the abnormality occurrence section from all sections of the time series with respect to the processing target data acquired by the data acquiring unit, and generates simulated abnormal data that simulates the abnormal data by multiplying each sample value of the physical quantity belonging to the processing section by the modulation pattern; A state determination device comprising:

2. the modulation pattern is obtained by dividing each sample value of the physical quantity belonging to the abnormality occurrence section by a representative value of the physical quantity belonging to a normal section of the entire time series, excluding the abnormality occurrence section. The state determination device according to claim 1 .

3. The abnormality occurrence section includes a transition section in which the state of the monitored object transitions from normal to abnormal. The state determination device according to claim 1 .

4. The processing section is a section in which the physical quantity is relatively large compared to other sections among all sections of the time series. The state determination device according to claim 1 .

5. The simulated abnormality data is used for a learning process or a verification process for an anomaly detection model that receives time-series data of the physical quantity as an input and outputs a determination result as to whether the state of the monitored object is normal or abnormal. The state determination device according to claim 1 .

6. a model evaluation unit that receives the time-series data of the physical quantity as an input, calculates an evaluation index related to the detection performance of an anomaly detection model that outputs a determination result as to whether the state of the monitored object is normal or abnormal, and evaluates the anomaly detection model based on the evaluation index; The state determination device according to claim 1 .

7. The evaluation index includes the number of erroneous determinations, The number of erroneous determinations is the number of times that an abnormality is determined to exist within a determination period that does not include the abnormality occurrence section, with respect to a time transition of the determination result obtained by inputting time series data of the physical quantity into the abnormality detection model. The state determination device according to claim 6.

8. The evaluation index includes a grace period, the grace period is, with respect to a time transition of the determination result obtained by inputting time series data of the physical quantity into the anomaly detection model, a time from a point in time when an anomaly is first determined to exist within a determination period including the anomaly occurrence section to a point in time when an anomaly occurs that is included in the time series data of the physical quantity. The state determination device according to claim 7.

9. the model evaluation unit calculates the number of times of erroneous determination and the grace period corresponding to a plurality of model parameter candidates for identifying a model structure of the anomaly detection model, and selects one model parameter candidate from the plurality of model parameter candidates according to a combination of the number of times of erroneous determination and the grace period. The state determination device according to claim 8 .

10. the model evaluation unit selects the one model parameter candidate for which the number of erroneous determinations is below a threshold and the grace period is the longest. The state determination device according to claim 9.

11. a learning processing unit that performs a learning process on the anomaly detection model using the anomaly data acquired by the data acquisition unit or the simulated anomaly data generated by the data generation unit, The state determination device according to claim 6.

12. the monitored object is a rotating machine installed in a facility, the model evaluation unit provisionally determines the anomaly detection model for each of the facilities or each of the types of the rotating machines, the learning processing unit performs additional or re-learning processing on the anomaly detection model provisionally determined by the model evaluation unit, thereby finally determining the anomaly detection model to be applied to each of the rotating machines. The state determination device according to claim 11.

13. the anomaly detection model determines whether a state of the monitored object is normal or abnormal based on a magnitude relationship between an error index related to a reconstruction error when the time-series feature quantities related to the physical quantities are reconstructed through dimensionality compression and dimensionality restoration, and a threshold value corresponding to the error index; the monitored object is a rotating machine, the physical quantity includes a vibration velocity, a vibration acceleration, and a temperature of the rotating machine; The state determination device according to claim 1 .

14. the error index is a linear sum of a first error, which is a standardized or normalized reconstruction error of the vibration velocity, a second error, which is a standardized or normalized reconstruction error of the vibration acceleration, and a third error, which is a standardized or normalized reconstruction error of the temperature; a weighting coefficient of the third error included in the linear sum is smaller than weighting coefficients of the first error and the second error included in the linear sum; The state determination device according to claim 13.

15. the anomaly detection model includes an unsupervised learner that reconstructs the time-series feature quantity through dimensionality reduction and dimensionality restoration of the time-series feature quantity related to the physical quantity; the time-series feature amount includes a plurality of partial time-series data extracted by sequentially sliding an extraction interval starting from a reference time point from the time-series data of the physical quantity; The state determination device according to claim 5 .

16. an acquisition step of acquiring abnormality data, which is time-series data of physical quantities correlated with vibrations of a monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the state of the monitored object within the entire section of the time series, and processing target data, which is time-series data of the physical quantities and is an object to be processed; a calculation step of calculating a modulation pattern indicating a time change in the relative intensity of the vibration in the abnormality occurrence section using the acquired abnormality data; a generating step of extracting a processing section equal to the length of the abnormality occurrence section from all sections of the time series for the acquired processing target data, and multiplying each sample value of the physical quantity belonging to the processing section by the modulation pattern to generate simulated abnormal data that simulates the abnormal data; A state determination program that causes one or more computers to execute the above.

17. a monitored object exhibiting vibration-related behavior; a sensor that measures a physical quantity correlated with vibration of the monitored object and outputs a measurement signal; a state determination device that acquires the measurement signal from the sensor and outputs a determination result as to whether the state of the monitored object is normal or abnormal; Equipped with The state determination device is a data acquisition unit that acquires abnormality data, which is time-series data of physical quantities correlated with vibrations of a monitored object and includes an abnormality occurrence section indicating that an abnormality has occurred in the state of the monitored object within the entire section of the time series, and processing target data, which is time-series data of the physical quantities and is an object to be processed; a pattern calculation unit that calculates a modulation pattern indicating a time change in the relative intensity of the vibration in the abnormality occurrence section using the abnormality data acquired by the data acquisition unit; and a data generating unit that extracts a processing section equal to the length of the abnormality occurrence section from all sections of the time series with respect to the processing target data acquired by the data acquiring unit, and generates simulated abnormal data that simulates the abnormal data by multiplying each sample value of the physical quantity belonging to the processing section by the modulation pattern; A state determination system comprising:

Images