State analysis apparatus, state analysis program, and state analysis system

Abstract

To provide a state analysis apparatus, a state analysis program, and a state analysis system which can specify cause of occurrence of an event with high accuracy.SOLUTION: A state analysis apparatus (20) calculates reconfiguration error 84 caused when a state amount set 80 is reconfigured through dimensional compression and dimensional restoration of the state amount set 80 as an aggregate of state amounts of tags at the same or corresponding time, calculates, among the entire period (B1) of time-serial data, as to each tag, error integrated value as an integrated value of reconfiguration error in interest periods (B1, B2) including the occurrence point of an event, and selects, based on an integrated value set 86 as an aggregate of the error integrated values of respective tags, a cause tag candidate as a cause tag or a candidate of the cause tag indicative of the occurrence cause of the event.SELECTED DRAWING: Figure 4

Description

[Technical Field]

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

[0002] 2. Description of the Related Art Conventionally, a condition analysis system has been known that sequentially measures a condition quantity indicating a condition within a facility, performs an analytical process on the obtained time-series data, and analyzes the condition within the facility.

[0003] Patent Document 1 discloses a method for generating diagnostic data from vibration waveform data of a target facility, determining signs of anomalies based on the distribution of reconstruction errors of the diagnostic data, and estimating the cause of an anomaly in the target facility based on an evaluation of errors in the diagnostic data for each frequency. Patent Document 1 also describes that when estimating the cause of an anomaly, a relationship diagram that defines the cause of an anomaly for each frequency range of the target facility is referenced.

[0004] Patent Document 2 discloses an operation management system in which operation data from multiple target locations is collected at a central location to detect signs of abnormality, with part of the detection processing being performed at each location. Patent Document 2 also describes that when an abnormality detection alert is generated, operation data including collected items that are effective in identifying the cause of the abnormality are transferred to the central location as data for analyzing the cause of the abnormality. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2020-030111 [Patent Document 2] Patent No. 6438124 Summary of the Invention [Problem to be solved by the invention]

[0006] Generally speaking, when the frequency of an event is relatively low, event detection models often use "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 many examples of an event occurring; (2) the conditions under which an event occurs are diverse, and unknown events may not be detected well; or (3) it is difficult to strictly define the occurrence of an event.

[0007] Furthermore, the likelihood of an event occurring does not necessarily correlate with the magnitude of the reconstruction error. This is because, for example, a phenomenon occurring upstream in a process may be delayed and amplified as it moves downstream. This may lead to, for example, a process monitor misidentifying the source of the event, which could delay response to the event.

[0008] The present disclosure has been made in consideration of the above-mentioned problems, and its purpose is to provide a status analysis device, a status analysis program, and a status analysis system that can identify the cause of an event with higher accuracy. [Means for solving the problem]

[0009] In order to solve the above problem, a state analysis device according to one aspect of the present disclosure includes a data acquisition unit that acquires a plurality of time series data indicating the time progression of state quantities that indicate a state within a facility by associating the data with tags; an error calculation unit that calculates a reconstruction error when a state quantity set, which is a collection of the state quantities for each tag at the same or corresponding time point, is reconstructed through dimensionality reduction and dimensionality restoration of the state quantity set, and calculates, for each tag, an error accumulation value that is an accumulation value of the reconstruction error within a section of interest that includes the time point at which an event occurs within the entire section of the time series data; and a tag selection unit that performs a selection process to select a cause tag that indicates the cause of the event or a candidate cause tag that is a candidate for the cause tag from two or more tags, based on the accumulation value set that is a collection of the error accumulation values ​​for each tag calculated by the error calculation unit.

[0010] In another aspect of the present disclosure, a state analysis program causes one or more computers to execute the following steps: an acquisition step of acquiring, in association with tags, multiple time series data indicating the time progression of state quantities indicating the state within a facility; a calculation step of calculating a reconstruction error when the state quantity set, which is a collection of the state quantities for each tag at the same or corresponding time point, is reconstructed through dimensionality compression and dimensionality restoration of the state quantity set; and a selection step of performing a selection process of selecting, from two or more tags, a cause tag indicating the cause of the occurrence of the event or a candidate cause tag that is a candidate for the cause tag, based on the calculated accumulation value set, which is a collection of the error accumulation values ​​for each tag.

[0011] In another aspect of the present disclosure, a condition analysis system includes a monitored object located within a facility, a plurality of sensors that measure state quantities indicating the state of the monitored object and output measurement signals, and a condition analysis device that acquires each of the measurement signals from the plurality of sensors and analyzes the state of the monitored object, wherein the condition analysis device includes a data acquisition unit that acquires a plurality of time series data indicating the time progression of the state quantities in association with tags, an error calculation unit that calculates a reconstruction error when the state quantity set, which is a collection of the state quantities for each of the tags at the same or corresponding time points, is reconstructed through dimensionality reduction and dimensionality restoration of the state quantity set, and calculates, for each of the tags, an error accumulation value that is an accumulation value of the reconstruction error within a section of interest that includes the time point of the event out of the entire section of the time series data, and a tag selection unit that performs a selection process to select a cause tag indicating the cause of the occurrence of the event or a cause tag candidate that is a candidate for the cause tag from two or more of the tags based on the accumulation value set, which is a collection of the error accumulation values ​​for each of the tags calculated by the error calculation unit. [Effects of the Invention]

[0012] According to the present disclosure, the cause of an event can be identified with higher accuracy. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is an overall configuration diagram of a condition analysis 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 analysis device in FIG. 1. FIG. [Figure 3] FIG. 3 is a diagram schematically illustrating an example of the time-series data group of FIG. 2. [Figure 4] FIG. 3 is a functional block diagram of an error calculation unit shown in FIG. 2. [Figure 5] FIG. 10 is a diagram showing an example of time changes of two state quantities in an abnormality occurrence section. [Figure 6] 3 is a flowchart showing an example of an analysis operation by the state analysis device of FIGS. 1 and 2. [Figure 7] FIG. 10 is a diagram illustrating an example of time changes in state quantities and reconstruction errors within an abnormality occurrence section. [Figure 8] 3 is a diagram schematically illustrating input / output characteristics of the event detection model and the analysis model of FIG. 2. FIG. [Figure 9] 3 is a functional block diagram of a third selection unit shown in FIG. 2. FIG. [Figure 10] FIG. 10 is a diagram showing an example of an evaluation result based on the explanatory model of FIG. 9. DETAILED DESCRIPTION OF THE INVENTION

[0014] First, some aspects of the present disclosure will be described.

[0015] A state analysis device according to a first aspect of the present disclosure includes a data acquisition unit that acquires a plurality of time series data indicating time progression of state quantities indicating a state within a facility in association with tags; an error calculation unit that calculates a reconstruction error when a state quantity set, which is a collection of the state quantities for each tag at the same or corresponding time point, is reconstructed through dimensionality compression and dimensionality restoration of the state quantity set, and calculates, for each tag, an error accumulation value that is an accumulation value of the reconstruction error within a section of interest that includes the time point at which an event occurs out of the entire section of the time series data; and a tag selection unit that performs a selection process to select a cause tag indicating a cause of the occurrence of the event or a candidate cause tag that is a candidate for the cause tag from two or more tags, based on the accumulation value set that is a collection of the error accumulation values ​​for each tag calculated by the error calculation unit.

[0016] In the condition analysis device of the second aspect of the present disclosure, the selection process may include a comparison operation that compares the error accumulation value with a threshold value, or compares the error accumulation values ​​with each other to select the cause tag or the candidate cause tag.

[0017] In the state analysis device of the third aspect of the present disclosure, the selection process may include a comparison operation in which a state quantity integrated value, which is an integrated value of the state quantity within the section of interest, is calculated for each tag, an integrated value ratio, which is the ratio of the error integrated value to the state quantity integrated value, is calculated for each tag, and the integrated value ratio is compared with a threshold value, or the integrated value ratios are compared with each other to select the cause tag or the cause tag candidate.

[0018] In the state analysis device according to the fourth aspect of the present disclosure, the selection process may include a comparison operation in which the contribution of the state quantity to the error accumulation value is calculated for each tag, and the contribution is compared with a threshold value, or the contributions are compared with each other to select the cause tag or the candidate cause tag.

[0019] In the state analysis device according to the fifth aspect of the present disclosure, the contribution may be determined using an explanatory model for explaining the contribution of an input value to a computational model that takes an error set, which is a collection of reconstruction errors for each tag, as input and outputs the accumulated value set or an error feature generated from the accumulated value set.

[0020] In the state analysis device according to a sixth aspect of the present disclosure, the number of the tags constituting the error set may be equal to or less than the number of the tags constituting the state quantity set.

[0021] In the state analysis device according to a seventh aspect of the present disclosure, the selection process may include a comparison operation for calculating a correlation coefficient between the state quantity and the reconstruction error within the section of interest for each tag, and comparing the correlation coefficient with a threshold value, or comparing the correlation coefficients with each other to select the cause tag or the candidate cause tag.

[0022] In the state analysis device according to an eighth aspect of the present disclosure, the selection process includes a first comparison operation that calculates the contribution of the reconstruction error to the error accumulation value for each tag, and compares the contribution with a threshold corresponding to the contribution, or compares the contributions with each other to select the cause tag or the cause tag candidate, and a second comparison operation that calculates a correlation coefficient between the state quantity and the reconstruction error in the section of interest for each tag, and compares the correlation coefficient with a threshold corresponding to the correlation coefficient, or compares the correlation coefficients with each other to select the cause tag or the cause tag candidate, and the tag selection unit may select the cause tag by executing the first comparison operation and the second comparison operation.

[0023] In the state analysis device according to a ninth aspect of the present disclosure, the event may be detection of an abnormality occurring within the facility or a suspicion of an abnormality within the facility.

[0024] A state analysis program in a tenth aspect of the present disclosure causes one or more computers to execute the following steps: an acquisition step of acquiring, in association with tags, multiple time series data indicating the time progression of state quantities indicating the state within a facility; a calculation step of calculating a reconstruction error when the state quantity set, which is a collection of the state quantities for each tag at the same or corresponding time point, is reconstructed through dimensionality compression and dimensionality restoration of the state quantity set; and a selection step of performing a selection process of selecting, from two or more tags, a cause tag indicating the cause of the occurrence of the event or a candidate cause tag that is a candidate for the cause tag, based on the calculated accumulation value set, which is a collection of the error accumulation values ​​for each tag.

[0025] A condition analysis system in an eleventh aspect of the present disclosure comprises a monitored object located within a facility, a plurality of sensors that measure state quantities indicating the state of the monitored object and output measurement signals, and a condition analysis device that acquires the measurement signals from the plurality of sensors and analyzes the state of the monitored object, wherein the condition analysis device comprises a data acquisition unit that acquires a plurality of time series data indicating time progression of the state quantities in association with tags, an error calculation unit that calculates a reconstruction error when the state quantity set, which is a collection of the state quantities for each of the tags at the same or corresponding time points, is reconstructed through dimensionality reduction and dimensionality restoration of the state quantity set, and calculates, for each of the tags, an error accumulation value that is an accumulation value of the reconstruction error within a section of interest that includes the time point of the event out of the entire section of the time series data, and a tag selection unit that performs a selection process to select a cause tag indicating the cause of the occurrence of the event or a cause tag candidate that is a candidate for the cause tag from two or more of the tags, based on the accumulation value set, which is a collection of the error accumulation values ​​for each of the tags calculated by the error calculation unit.

[0026] 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.

[0027] [Configuration of Condition Analysis System 10] <Overall structure> FIG. 1 is an overall configuration diagram of a condition analysis system 10 according to an embodiment of the present disclosure. The condition analysis system 10 is provided to analyze the condition of a monitored object. In the example of FIG. 1, the monitored object is provided in a facility that refines crude oil to produce and store petroleum products, i.e., an oil refinery 12. Specifically, the condition analysis system 10 includes equipment 14, devices 15, and piping 16 (corresponding to "monitored objects") within the oil refinery 12, a process monitoring device 18, and a condition analysis device 20. Here, the process monitoring device 18 and the condition analysis device 20 are configured to be able to communicate with each other via a network NT.

[0028] Facility 14 may be composed of, for example, a receiving facility, a crude oil tank, a refinery, or a product tank. Equipment 15 constitutes part of facility 14 or is installed in facility 14. Piping 16 is provided for transporting [1] materials to be processed, such as crude oil, intermediate products, and petroleum products, or [2] fluids including refrigerants and hydraulic oils.

[0029] One or more sensors 22 for measuring state quantities are attached to the facility 14, the equipment 15, or the piping 16. Each sensor 22 outputs a measurement signal indicating a time series of the state quantity to the process monitoring device 18. Examples of the state quantity include: [1] physical quantities related to the object to be processed, including temperature, pressure, flow rate, flow velocity, and viscosity; [2] physical quantities related to the operating state of the facility 14 or the equipment 15; or [3] physical quantities related to the external environment.

[0030] The process monitoring device 18 is a computer for monitoring the work processes within the refinery 12. The process monitoring device 18 receives, for example, measurement signals from the sensors 22 and transmits the obtained time-series data to the condition analysis device 20 periodically or irregularly. Prior to transmitting the time-series data, the process monitoring device 18 assigns unique identification information (hereinafter referred to as a "tag") to the data generation source or the sensor 22.

[0031] The process monitoring device 18 also receives the analysis results from the condition analysis device 20 and outputs the analysis results to an output device (not shown). This allows the operator to quickly understand that an event has occurred in a specific facility 14, equipment 15, or piping 16 within the refinery 12. In the example of FIG. 1 , an "event" refers to the detection of an abnormality that has occurred within the refinery 12 or a suspicion of an abnormality within the refinery 12. Note that the "time at which an event occurs" may also include, for example, the time at which an abnormality is actually confirmed through an inspection or the like of the facility 14, etc. installed within the refinery 12, the time at which an abnormality actually confirmed through an inspection or the like is suspected to have occurred, the time at which the process monitoring device 18 or the condition analysis device 20 issues an alert based on preset monitoring criteria in accordance with the acquired time-series data, the time at which an operator detects an abnormal sound or odor during a patrol inspection within the refinery 12, the time at which an operator suspects an abnormality based on the acquired time-series data, and the time at which a preset periodic inspection of the facility 14, etc. arrives.

[0032] The condition analysis device 20 performs various information processes to analyze the condition of the refinery 12 in response to a request made via the process monitoring device 18 after an operator recognizes the occurrence of an event, for example. Alternatively, the condition analysis device 20 performs various information processes to automatically analyze the condition of the refinery 12, either periodically or irregularly. The condition analysis device 20 is a server computer that performs overall control over the condition analysis of the refinery 12, and may be either a cloud-based or on-premise type. Here, the condition analysis device 20 is illustrated as a single computer, but the condition analysis device 20 may instead be a group of computers that form a distributed system.

[0033] <Configuration of the state analyzer 20> Fig. 2 is a block diagram showing an example of the configuration of the state analysis device 20 appearing in Fig. 1. Specifically, the state analysis device 20 includes a communication unit 30, a control unit 32, and a storage unit .

[0034] The communication unit 30 is an interface for transmitting and receiving electrical signals to and from external devices, which allows the condition analysis device 20 to acquire time-series data obtained through measurements by the sensors 22 from the process monitoring devices 18, and to supply the analysis result information 64 generated by itself to the process monitoring devices 18 in the corresponding refineries 12.

[0035] 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 learning processing unit 42, an error calculation unit 44, a tag selection unit 46, and an output instruction unit 48.

[0036] The data acquisition unit 40 acquires various data related to the learning process or the analysis process, such as a time-series data group 56 or tag information 58. The data acquisition unit 40 may acquire data via communication from an external device including the process monitoring device 18, or may acquire data by reading it from the storage unit 34.

[0037] The learning processing unit 42 performs a learning process on the anomaly detection model DM (corresponding to an "event detection model") or the analysis model AM using the time-series data group 56 acquired by the data acquiring unit 40. The type of learning process may be any of supervised learning, unsupervised learning, and reinforcement learning. Examples of learning parameter update rules include the stochastic gradient descent method, the momentum method, the AdaGrad method, and the Adam method. Through this learning process, each value of the first learning parameter group 60 or each value of the second learning parameter group 62 is determined.

[0038] The error calculation unit 44 calculates the reconstruction error for a collection of state quantities for each tag at the same or corresponding time points (hereinafter referred to as a state quantity set). Here, the "reconstruction error" refers to the error that occurs before and after reconstruction when reconstructing a state quantity set through dimensionality reduction and dimensionality restoration of the state quantity set. This reconstruction error is calculated for each element of the state quantity set (i.e., for each tag). Furthermore, the above "corresponding time points" means that the time points associated with each state quantity are the same within an allowable range (for example, within one day, within one hour, etc.).

[0039] The error calculation unit 44 calculates, for each tag, an integrated value of the reconstruction error (hereinafter also referred to as "integrated error value") within a section of interest that includes the time point at which an event occurs, out of all sections of the time series data. Here, the "section of interest" is a section that includes the time point at which the event occurs, and corresponds to the time range over which the reconstruction error is integrated. The section of interest may be an "occurrence section" extracted from before and after the time point at which the event occurs, or may be "the entire section" of the time series data.

[0040] The error calculation unit 44 is configured to include a plurality of computational models (or mathematical models), specifically, an anomaly detection model DM and an analysis model AM. The specific configuration of the anomaly detection model DM will be described in detail with reference to FIG. 4. The specific configuration of the analysis model AM will be described in detail with reference to FIG. 8.

[0041] The tag selection unit 46 performs a "selection process" to select a tag indicating the cause of an event (hereinafter referred to as a "cause tag") or a candidate cause tag (hereinafter referred to as a "candidate cause tag") from two or more tags. Specifically, the tag selection unit 46 includes a first selection unit 51, a second selection unit 52, a third selection unit 53, and a fourth selection unit 54.

[0042] The first selection unit 51 uses the error accumulation value calculated by the error calculation unit 44 (here, the anomaly detection model DM) to select two or more cause tag candidates (hereinafter also referred to as the "first candidate group") from all tags to be selected. Specifically, the first selection unit 51 calculates the error accumulation value for each tag, and then performs a "comparison operation" in which the error accumulation value is compared with a threshold value or the error accumulation values ​​are compared with each other to select cause tag candidates. The accumulation range of the error accumulation value (i.e., the interval of interest) is set to, for example, the entire interval of the time-series data. Note that the threshold value or the number / proportion of cause tag candidates used in the comparison operation can be set to various values.

[0043] The second selection unit 52 uses the error accumulation value calculated by the error calculation unit 44 (here, the anomaly detection model DM) to select two or more cause tag candidates (hereinafter also referred to as the "second candidate group") from the cause tag candidates belonging to the first candidate group. Specifically, the second selection unit 52 performs a "comparison operation" in which, after calculating the accumulation value ratio for each tag, the second selection unit 52 compares the accumulation value ratio with a threshold value or compares the accumulation value ratios with each other to select the cause tag candidates. This accumulation value ratio means the ratio of the error accumulation value to the accumulation value of the state quantity within the interval of interest (hereinafter referred to as the "state quantity accumulation value"). The accumulation range of the error accumulation value (i.e., the interval of interest) is set, for example, to the interval in which the event occurs. Note that the threshold value or the number / proportion of cause tag candidates used in the comparison operation can be set to various values.

[0044] The third selection unit 53 selects two or more cause tag candidates (hereinafter also referred to as the "third candidate group") from the cause tag candidates belonging to the second candidate group based on the error accumulation value calculated by the error calculation unit 44 (here, the anomaly detection model DM). Specifically, the third selection unit 53 calculates the contribution of each tag, and then performs a "comparison operation" (corresponding to the "first comparison operation") by comparing the contribution with a threshold or by comparing the contributions with each other to select a cause tag or a cause tag candidate. The accumulation range of the error accumulation value (i.e., the interval of interest) is set, for example, to the interval in which the event occurs. This contribution is a quantitative value related to the degree to which the state quantity contributes to the error accumulation value, and is calculated using the analysis model AM and the explanation model EM. The method for calculating the contribution by the third selection unit 53 will be described in detail with reference to FIG. 9. The specific configuration of the explanation model EM will be described in detail with reference to FIG. 9. The threshold or the number and proportion of cause tag candidates used in the comparison operation can be set to various values.

[0045] The fourth selection unit 54 selects one or more cause tag candidates from the cause tag candidates belonging to the third candidate group and identifies them as cause tags. Specifically, the fourth selection unit 54 calculates the correlation coefficient for each tag, and then performs a "comparison operation" (corresponding to the "second comparison operation") in which the correlation coefficient is compared with a threshold value or the correlation coefficients are compared with each other to select the cause tag. This correlation coefficient is calculated between the state quantity and the error accumulation value within the interval of interest. The accumulation range of the correlation coefficient (i.e., the interval of interest) is set, for example, to the interval in which the event occurs. Note that the threshold value or the number / proportion of cause tags used in the comparison operation can be set to various values. For example, the number of cause tags is preferably a value between 2 and 5 from the perspective of suggesting it to the operator.

[0046] The output instruction unit 48 instructs the output device to present the analysis results obtained by the tag selection unit 46 to the operator. This "output device" may be a device provided in an external device including the process monitoring device 18, or may be an output device (not shown) provided in the condition analysis device 20. The output may be in any form that stimulates the five senses, such as sight, hearing, and touch.

[0047] The storage unit 34 stores programs and data necessary for the control unit 32 to control each component. The storage unit 34 is composed of a non-transitory, computer-readable recording medium. Here, the computer-readable recording 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.

[0048] In the example of FIG. 2, the storage unit 34 stores a time-series data group 56, tag information 58, a first learning parameter group 60, a second learning parameter group 62, and analysis result information 64.

[0049] The time series data group 56 is a collection of time series data used for learning processing or analysis processing. The time series data is data showing the time transition of a state quantity, and is a data set including data pairs of a state quantity and a measurement time point. The sampling periods of the time series data are all shaped to be equal.

[0050] The tag information 58 includes, for example, [1] tag identification information, [2] meta information about the sensor 22, [3] meta information about the data source (here, equipment 14, device 15, piping 16), or [4] the name of the facility (here, refinery 12).

[0051] The first learning parameter group 60 is a collection of learning parameters identified through a learning process for the anomaly detection model DM. The first learning parameter group 60 is determined for each refinery 12 registered in the condition analyzing device 20, for example.

[0052] The second learning parameter group 62 is a collection of learning parameters identified through a learning process for the analytical model AM. Similar to the first learning parameter group 60, the second learning parameter group 62 is determined for each refinery 12 registered in the condition analyzing device 20.

[0053] Examples of learning parameters include [1] "model parameters" (or hyperparameters) that specify the model structure of a learning model, and [2] "variable parameters" whose optimal values ​​change depending on the population of training data. Examples of variable parameters include weight coefficients between arithmetic units, thresholds for activation functions, and principal component coefficients.

[0054] The analysis result information 64 includes [1] "time information" such as the date and time when the event occurred, [2] "analysis results" such as the type of event, identification information of the cause tag, accuracy of the analysis, urgency and priority of the response, or [3] "meta information" regarding the facility 14, equipment 15, piping 16 or refinery 12.

[0055] <Explanation of the anomaly detection model DM> Next, the abnormality detection model DM included in the error calculation unit 44 in FIG. 2 will be described with reference to FIGS.

[0056] FIG. 3 is a diagram schematically illustrating an example of the time-series data group 56 of FIG. 2. The horizontal axis of the graph indicates time (unit: days), and the vertical axis of the graph indicates state quantities (unit: arbitrary). In the example of FIG. 3, the time-series data group 56 is composed of N pieces of time-series data. N is an arbitrary natural number, and can take on values ​​ranging from tens to hundreds, for example. Here, the entire section B1 corresponds to the time range (here, t=T0 to T3) included in all of the N pieces of time-series data. Furthermore, the abnormality occurrence section B2 corresponds to the time range (here, t=T1 to T2) in which an abnormality occurred at any location within the refinery 12. Note that, since the N state quantities have different definitions, normalization or standardization is performed to match the scale of each data.

[0057] Fig. 4 is a functional block diagram related to the error calculation unit 44 shown in Fig. 2. The error calculation unit 44 is configured to include an anomaly detection model DM and an integration processing unit 70. The anomaly detection model DM also includes an unsupervised learner 72, a difference processing unit 74, a synthesis unit 76, and a threshold processing unit 78.

[0058] The unsupervised learner 72 performs a "reconstruction operation" to reconstruct the state quantity set through dimensionality reduction and dimensionality restoration of the state quantity set. Examples of reconstruction operations include principal component analysis or an autoencoder. An original state quantity set 80 is input through the unsupervised learner 72, and a reconstructed state quantity set 82 is output. The original state quantity set 80 is a data sequence in which N state quantities at the same or corresponding time points are arranged in a predetermined tag order. The reconstructed state quantity set 82 is a data sequence obtained by reconstructing the original state quantity set 80 through the unsupervised learner 72. Hereinafter, the time series of the state quantity set 82 within the interval of interest may be referred to as the "first set time series 80s."

[0059] The difference processing unit 74 performs a "difference calculation" to find the difference between each element of the state quantity sets 80 and 82 and calculate a reconstruction error for each tag (i.e., an error set 84). This reconstruction error may be an absolute error (AE) or the square of the error (RMS). The error set 84 is sequentially output to the integration processing unit 70.

[0060] The combiner 76 performs a "combining operation" to combine the reconstruction errors (i.e., error set 84) for each tag calculated by the differential processor 74 and calculate a combined error at the time of determination. This combined error may be either a mean absolute error (MAE) or a root mean square error (RMSE).

[0061] The threshold processing unit 78 performs a "threshold calculation" to output a flag value indicating whether the result of the determination is normal or abnormal, depending on the magnitude relationship between the combined error calculated by the combining unit 76 and the threshold value Th. This flag value is defined so that, for example, 0 indicates "normal" and 1 indicates "abnormal." This makes it possible to determine the time point when the abnormality occurred or the abnormality occurrence section B2 (FIG. 3).

[0062] The accumulation processing unit 70 accumulates the time series of the error set 84 in the section of interest (hereinafter, second set time series 84s) for each tag, with respect to the time series of the error set 84 sequentially output from the anomaly detection model DM. This results in a collection of error accumulation values ​​for each tag (hereinafter, accumulation value set 86).

[0063] [Operation of Condition Analysis System 10] The condition analyzing system 10 in this embodiment is configured as described above. Next, the operation of the condition analyzing system 10, particularly the analysis operation by the condition analyzing device 20, will be outlined with reference to FIGS.

[0064] <1.Problems> Figure 5 shows an example of the time changes of two state quantities within the abnormality occurrence section B2. The horizontal axis of the graph indicates time (unit: hour), and the vertical axis of the graph indicates the state quantity for each tag. The graph shown by the solid line corresponds to the "original state quantity," which is the input value of the unsupervised learning device 72 (Figure 3). The graph shown by the dashed line corresponds to the "reconstructed state quantity," which is the output value of the unsupervised learning device 72 (Figure 3).

[0065] In the example of FIG. 5, the "state quantity 3" on the upper side is the state quantity of the equipment 15 to which a "non-cause tag" (tag ID = 3) that is not the cause of the abnormality has been assigned (tag ID = 3) is attached. The "state quantity 4" on the lower side is the state quantity of the piping 16 to which a "cause tag" (tag ID = 4) that is the cause of the abnormality has been assigned (tag ID = 4). As can be seen from FIG. 5, within the abnormality occurrence section B2, the reconstruction error for the non-cause tag is relatively small, and the reconstruction error for the cause tag is relatively large. In other words, it cannot be said that the likelihood of an abnormality occurring is necessarily correlated with the magnitude of the reconstruction error.

[0066] The reason for this is that, for example, an abnormal phenomenon occurring upstream in the oil refining process may be delayed and amplified as it moves downstream. If an operator mistakenly identifies the device 15 with tag ID = 3 as the source of the abnormality, the operator may prioritize checking the operation of that device 15, which may delay the discovery of the abnormality in the piping 16. Therefore, the condition analysis device 20 in Figure 2 performs analysis processing to more accurately identify the cause tag even when the tendency illustrated in Figure 5 appears.

[0067] <2. Overall movement> FIG. 6 is a flowchart showing an example of the analysis operation by the state analysis device 20 of FIGS.

[0068] In step SP10, the data acquisition unit 40 of the state analysis device 20 acquires a collection of time-series data collected via the process monitoring device 18 in the refinery 12 (that is, the time-series data group 56).

[0069] In step SP12, the learning processing unit 42 of the state analysis device 20 performs learning processing on the anomaly detection model DM using the time-series data group 56 acquired in step SP10. As a result, a learned anomaly detection model DM is generated.

[0070] In step SP14, the error calculation unit 44 of the state analysis device 20 calculates an error integrated value (that is, integrated value set 86) for each tag using the abnormality detection model DM generated through the learning process in step SP12.

[0071] Fig. 7 is a diagram showing an example of changes over time in the state quantity and reconstruction error within the abnormality occurrence section B2. The horizontal axis of the graph indicates time (unit: hour), and the vertical axis of the graph indicates [1] the magnitude of the state quantity (unit: arbitrary) or [2] the reconstruction error (unit: dimensionless). In the example of Fig. 7, only the abnormality occurrence section B2 is partially shown out of the entire time series section B1. Here, the error integrated value corresponds to the area enclosed by the graph showing the reconstruction error and the horizontal axis (time axis).

[0072] Note that the abnormal occurrence section B2 is configured to include some or all of the abnormal sections. That is, the abnormal occurrence section B2 is any one of [1] only abnormal sections, [2] a combination of abnormal sections and normal sections, [3] a combination of transition sections and abnormal sections, or [4] a combination of normal sections, transition sections, and abnormal sections. Here, the "transition section" means a section until the state of the oil refinery 12 transitions from normal to abnormal.

[0073] In step SP16 of FIG. 6, the first selection unit 51 of the state analysis device 20 performs a "primary selection process" of extracting N1 (1 < N1 < N) cause tag candidates from among N tags using the integrated value set 86 calculated in step SP14.

[0074] In step SP18, the second selection unit 52 of the state analysis device 20 performs a "secondary selection process" of extracting N2 (1 < N2 < N1) cause tag candidates from among N1 tags using the integrated value set 86 calculated in step SP14.

[0075] Returning to FIG. 7, the state quantity and the reconstruction error each tend to transition with relatively large values while gradually increasing in the transition section until transitioning from normal to abnormal within the abnormal occurrence section B2. That is, within the abnormal occurrence section B2, the state quantity and the reconstruction error often have a relatively high correlation. Therefore, the second selection unit 52 may respectively obtain the correlation coefficients corresponding to the N1 tags and select N2 tags whose correlation coefficients are higher than the threshold value.

[0076] In step SP20, the learning processing unit 42 of the state analysis device 20 extracts data corresponding to the N2 cause tag candidates from the time-series data group 56 acquired in step SP10, and then performs learning processing on the analysis model AM. Thereby, a learned analysis model AM is generated.

[0077] FIG. 8 is a diagram schematically illustrating the input / output characteristics of the anomaly detection model DM and the analysis model AM of FIG. 2. As already described in FIG. 4, the anomaly detection model DM is a computation model that receives a state quantity set 80 consisting of N elements as input and outputs an error set 84 consisting of N elements. A second set of time series 84s is generated from a first set of time series 80s through sequential input / output. Then, an integrated value set 86 is generated from the second set of time series 84s through integration by the integration processing unit 70. Furthermore, the error calculation unit 44 generates a representative integrated value 94 from the integrated value set 86. The representative integrated value 94 is one representative value (e.g., a statistical value such as a maximum value, minimum value, or median) of the N elements constituting the integrated value set 86, but may be two or more representative values. Furthermore, the representative integrated value 94 may be the integrated values ​​themselves constituting the integrated value set 86, or may be a feature (hereinafter, referred to as an error feature) generated from the integrated value set 86. An example of the error feature is an average value or a weighted sum of a plurality of integrated values.

[0078] Through the primary and secondary selection processes, the number of tags is selected from N to N2. As a result of this selection, error set 84 consisting of N elements is divided into error subset 90 consisting of N2 elements and error subset 92 consisting of (N-N2) elements. Hereinafter, the collections of error subsets 90 and 92 within the interval of interest are referred to as subset time series 90s and 92s.

[0079] The analytical model AM is a regression model that uses the error subset 90 at each time point constituting the subset time series 90s as an explanatory variable and the representative integrated value 94 as a response variable. For example, when the analytical model AM receives the error subset 90 as an input, it can output the representative integrated value 94. Examples of regression models include linear regression, ridge regression, lasso regression, elastic net regression, logistic regression, random forest, gradient boosting decision tree, support vector machine, and neural network regression.

[0080] In step SP22 of FIG. 6, the third selection unit 53 of the state analysis device 20 performs a "third selection process" of extracting N3 (1 < N3 < N2) cause tag candidates from among N2 tags using the analysis model AM generated in step SP20.

[0081] FIG. 9 is a functional block diagram of the third selection unit 53 shown in FIG. 2. The third selection unit 53 is configured to include an explanatory model EM for explaining the contribution degree of the input value to the analysis model AM with respect to the representative integrated value 94. The explanatory model EM calculates the contribution degree for each cause tag candidate using the subset time series 90s and the representative integrated value 94.

[0082] The explanatory model EM is classified into a "local explanatory model" for visualizing the basis for judging each data and a "global explanatory model" for visualizing the interpretation of the model for the entire data. Here, for the purpose of selecting cause tag candidates, the local explanatory model is used for the explanatory model EM. Examples of the local explanatory model include LIME (Local Interpretable Model-agnostic Explanation), SHAP (SHapley Additive exPlanations), Anchor, or Influence.

[0083] FIG. 10 is a diagram showing an example of the evaluation result by the explanatory model EM of FIG. 9. The vertical axis of the bar graph (that is, the arrangement of the bars) indicates the tag name, and the horizontal axis of the bar graph (that is, the length of the bars) indicates the SHAP value, respectively. The SHAP value is defined such that the larger the value, the higher the contribution degree, and the smaller the value, the lower the contribution degree. In the example of FIG. 10, the tags with larger SHAP values are located on the upper side, that is, the tag names are displayed in descending order of the SHAP value. The second tag "519" surrounded by the broken line corresponds to the actual cause tag. That is, the evaluation index (that is, the contribution degree) using the SHAP value indicates that it is effective for specifying the cause tag.

[0084] In step SP24 of FIG. 6, the fourth selection unit 54 of the state analysis device 20 performs a "fourth selection process" for identifying N4 (1≦N4<N3) cause tags from among the N3 tags selected in step SP22. Specifically, the fourth selection unit 54 obtains a correlation coefficient for each of the N3 tags, and identifies N4 tags with relatively or absolutely large correlation coefficients as "cause tags". Through this series of selection processes, analysis result information 64 including information on the identified cause tags is generated.

[0085] In step SP26, the output instruction unit 48 of the state analysis device 20 transmits data including the analysis result information 64 obtained in step SP24. As a result, the analysis result information 64 is output via the process monitoring device 18.

[0086] In this way, the state analysis device 20 finishes the analysis operation on the time-series data group 56 according to the flowchart of FIG. 6. As a result, the cause of the abnormality in the oil refinery 12 is identified with higher accuracy.

[0087] [Summary of the Embodiment] As described above, the state analysis device 20 in this embodiment calculates, for each tag, an error integrated value that is the integrated value of the reconstruction errors within the region of interest (here, the entire region B1 or the abnormal occurrence region B2) including the event occurrence time point among all time-series data, and based on the integrated value set 86, performs a selection process for selecting a cause tag or a cause tag candidate from among two or more tags. By using the error integrated value with a high correlation with the cause of the event, the cause of the event can be identified with higher accuracy.

[0088] Also, the selection process may include a comparison operation for comparing the error integrated value with a threshold value or comparing error integrated values with each other to select a cause tag or a cause tag candidate. By comparing the magnitude relationship of the error integrated values themselves, the selection accuracy of the cause tag or the cause tag candidate is further improved.

[0089] Further, the selection process may include a comparison operation of obtaining an integrated value ratio, which is the ratio of the error integrated value to the state quantity integrated value, for each tag, comparing the integrated value ratio with a threshold value, or comparing the integrated value ratios with each other to select a cause tag or a cause tag candidate. By using the integrated value ratio obtained by dividing the error integrated value indicating the noise amount by the state quantity integrated value indicating the signal amount, the selection accuracy of the cause tag or the cause tag candidate is further improved.

[0090] Further, the selection process may include a comparison operation of obtaining the contribution degree of the reconstruction error to the error integrated value for each tag, comparing the contribution degree with a threshold value, or comparing the contribution degrees with each other to select a cause tag or a cause tag candidate. By using the contribution degree, the selection accuracy of the cause tag or the cause tag candidate is further improved.

[0091] Further, the contribution degree may be obtained by using an explanation model EM for explaining the contribution degree of the input value to the operation model to the output value from the operation model, with respect to an operation model that takes the error subset 90 as an input and outputs the integrated value set 86 or an error feature amount generated from the integrated value set 86. By using the explanation model EM, the selection accuracy of the cause tag or the cause tag candidate is further improved.

[0092] Further, the number (N2) of tags constituting the error subset 90 may be equal to the number (N) of tags constituting the state quantity set 80, or may be less than N. In particular, when N2 < N is satisfied, the amount of calculation and calculation time by the explanation model EM can be significantly reduced through the reduction of the number of tags, which is the number of input values.

[0093] Further, the selection process may include a comparison operation of obtaining the correlation coefficient between the state quantity and the reconstruction error within the region of interest for each tag, comparing the correlation coefficient with a threshold value, or comparing the correlation coefficients with each other to select a cause tag or a cause tag candidate. By using the correlation coefficient, the selection accuracy of the cause tag or the cause tag candidate is further improved. <000​Furthermore, when the selection process includes a first comparison operation that selects a cause tag or a cause tag candidate using the contribution degree and a second comparison operation that selects a cause tag or a cause tag candidate using the correlation coefficient, the tag selector 46 may select a cause tag by sequentially executing the first comparison operation and the second comparison operation. Screening using the error accumulation value can prevent cause tags from being overlooked, thereby improving the accuracy of identifying the cause tag.

[0095] The event may also be the detection of an abnormality occurring within a facility (here, refinery 12) or a suspected abnormality within the facility. Through the identification of the selected cause tag, the abnormal condition within refinery 12 can be quickly addressed.

[0096] [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 the flowchart may be changed within the scope of no technical contradiction.

[0097] In the above embodiment, an example was described in which the tag selector 46 (FIG. 2) identifies the cause tag through four selection stages, but the method for identifying the cause tag is not limited to this. The number of selection stages may be one to three, or five or more. Furthermore, when selecting tags, other indices or combinations of these indices may be used instead of the above-mentioned error accumulation value, accumulation value ratio, contribution rate, or correlation coefficient.

[0098] In the above embodiment, the facility is an oil refinery 12, but the type of facility is not limited to this. For example, the facility may be a factory for producing various industrial products, or an outdoor location other than a factory, such as a chemical factory, steel mill, or power plant. Furthermore, the type of event is not limited to an "abnormality" in the production process, and may be, for example, congestion, waiting, inventory shortage, labor shortage, etc.

[0099] In the above embodiment, the process monitoring device 18 and the condition analysis device 20 are described as being capable of communicating with each other via the network NT, but this is not limiting. For example, the process monitoring device 18 may be capable of one-way communication from the process monitoring device 18 to the condition analysis device 20 via the network NT. In this case, the condition analysis device 20 may transmit (output) the analysis results via the network NT to a terminal (not shown) owned by the user of the condition analysis system 10 or the operator of the refinery 12, for example.

[0100] In this specification or claims, information, physical quantities, feature quantities, sample values, indicators, parameters, etc. may be expressed using absolute values, may be expressed using relative values ​​from a predetermined value, or may be expressed using corresponding other information.

[0101] In this specification or claims, when expressions such as "obtaining / setting / using / based on 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 information as input" (including similar expressions), this includes cases where the result is obtained based solely on the information in question, 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 "outputs 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]

[0102] 10. Condition analysis system, 12. Refinery (facility), 14. Equipment (monitored object), 15. Device (monitored object), 16. Piping (monitored object), 18. Process monitoring device, 20. Condition analysis device, 22. Sensor, 40. Data acquisition unit, 42. Learning processing unit, 44. Error calculation unit, 46. Tag selection unit, 48. Output instruction unit, 56. Time series data group, 80, 82. State quantity set, 80s. First set time series, 84. Error set, 84s. Second set time series, 86. Accumulated value set, 90s, 92s. Subset time series, 94. Representative accumulated value, B1. All section (interval of interest), B2. Anomaly occurrence section (interval of interest), AM. Analysis model (computation model), DM. Anomaly detection model (event detection model), EM. Explanation model

Claims

1. a data acquisition unit that acquires a plurality of time-series data indicating time transitions of state quantities indicating states within the facility in association with tags; an error calculation unit that calculates a reconstruction error when a state quantity set, which is a collection of the state quantities for each tag at the same or corresponding time point, is reconstructed through dimensionality reduction and dimensionality restoration of the state quantity set, and calculates, for each tag, an error integrated value that is an integrated value of the reconstruction error within an interval of interest that includes a time point at which an event occurs, among all intervals of the time series data; a tag selection unit that performs a selection process to select a cause tag indicating a cause of the occurrence of the event or a candidate cause tag that is a candidate for the cause tag from two or more tags based on an integrated value set that is a collection of the error integrated values ​​for each tag calculated by the error calculation unit; A state analysis device comprising:

2. the selection process includes a comparison operation of comparing the error accumulation value with a threshold value or comparing the error accumulation values ​​with each other to select the cause tag or the cause tag candidate; The state analyzer according to claim 1 .

3. The selection process includes a comparison operation for calculating, for each tag, a state quantity integrated value, which is an integrated value of the state quantity within the section of interest, calculating, for each tag, an integrated value ratio, which is a ratio of the error integrated value to the state quantity integrated value, and comparing the integrated value ratio with a threshold value or comparing the integrated value ratios with each other to select the cause tag or the cause tag candidate. The state analyzer according to claim 1 .

4. the selection process includes a comparison operation of calculating a contribution of the reconstruction error to the error accumulation value for each tag, comparing the contribution with a threshold, or comparing the contributions with each other to select the cause tag or the cause tag candidate. The state analyzer according to claim 1 .

5. the contribution is calculated using an explanatory model that explains the contribution of an input value to a computation model with respect to an output value from the computation model, the explanatory model having an error set that is a collection of the reconstruction errors for each tag as an input and the integrated value set or an error feature amount generated from the integrated value set as an output. The state analyzer according to claim 4 .

6. the number of the tags constituting the error set is equal to or less than the number of the tags constituting the state quantity set; The state analyzer according to claim 5 .

7. the selection process includes a comparison operation of calculating a correlation coefficient between the state quantity and the reconstruction error within the section of interest for each tag, and comparing the correlation coefficient with a threshold or comparing the correlation coefficients with each other to select the cause tag or the cause tag candidate. The state analyzer according to claim 1 .

8. The selection process includes: a first comparison operation for calculating a contribution of the reconstruction error to the error accumulation value for each tag, and comparing the contribution with a threshold corresponding to the contribution, or comparing the contributions with each other to select the cause tag or the cause tag candidate; a second comparison operation for calculating a correlation coefficient between the state quantity and the reconstruction error within the section of interest for each tag, and comparing the correlation coefficient with a threshold corresponding to the correlation coefficient, or comparing the correlation coefficients with each other to select the cause tag or the cause tag candidate; the tag selection unit selects the cause tag by performing the first comparison operation and the second comparison operation; The state analyzer according to claim 1 .

9. The event is a detection of an abnormality occurring within the facility or a suspicion of an abnormality within the facility. The state analyzer according to claim 1 .

10. an acquisition step of acquiring a plurality of time-series data indicating time transitions of state quantities indicating states within the facility in association with tags; a calculation step of calculating a reconstruction error when a state quantity set, which is a collection of the state quantities for each tag at the same or corresponding time point, is reconstructed through dimensionality reduction and dimensionality restoration of the state quantity set, and calculating, for each tag, an error integrated value that is an integrated value of the reconstruction error within an interval of interest that includes a time point at which an event occurs, among all intervals of the time series data; a selection step of selecting a cause tag indicating a cause of the occurrence of the event or a candidate cause tag that is a candidate for the cause tag from two or more tags based on an accumulation value set that is a collection of the calculated error accumulation values ​​for each tag; A status analysis program that causes one or more computers to execute the above.

11. The monitored objects within the facility, a plurality of sensors that measure state quantities indicating the state of the object to be monitored and output measurement signals; a state analyzer that acquires the measurement signals from the plurality of sensors and analyzes the state of the monitored object; Equipped with The state analysis device a data acquisition unit that acquires a plurality of time-series data indicating a time transition of the state quantity in association with a tag; an error calculation unit that calculates a reconstruction error when a state quantity set, which is a collection of the state quantities for each tag at the same or corresponding time point, is reconstructed through dimensionality reduction and dimensionality restoration of the state quantity set, and calculates, for each tag, an error integrated value that is an integrated value of the reconstruction error within an interval of interest that includes a time point at which an event occurs, among all intervals of the time series data; a tag selection unit that performs a selection process to select a cause tag indicating a cause of the occurrence of the event or a candidate cause tag that is a candidate for the cause tag from two or more tags based on an integrated value set that is a collection of the error integrated values ​​for each tag calculated by the error calculation unit; A condition analysis system comprising:

Images