ABNORMALITY FACTOR ESTIMATION DEVICE, LEARNING DEVICE, PRECISE DIAGNOSIS SYSTEM, AND ABNORMALITY FACTOR ESTIMATION METHOD

Abstract

The present invention involves: a sensor data acquisition unit (10) that acquires a plurality of sets of time-series sensor data collected by a plurality of sensors (300) provided in a plurality of facility components constituting a target facility; an abnormality detection unit (30) that detects a plurality of abnormality-detecting sensors that are experiencing abnormality on the basis of the plurality of sets of sensor data; an abnormality detection order estimation unit (40) that estimates the abnormality detection order in which the plurality of abnormality-detecting sensors have been detected as experiencing abnormality; an abnormality propagation path tracing unit (50) that estimates the abnormality propagation order in which an abnormality has propagated on the basis of abnormality-detecting sensor information regarding the plurality of abnormality-detecting sensors and an estimated structure that shows dependencies between the facility components; and an abnormality factor estimation unit (60) that estimates factors for the abnormality on the basis of the abnormality detection order and the abnormality propagation order.

Description

[Technical field]

[0001] The present disclosure relates to an abnormality factor estimation device, a learning device, a precision diagnosis system, and an abnormality factor estimation method. [Background technology]

[0002] In facilities such as plants or factories, multiple devices and other elements that make up the facility (hereinafter referred to as "facility components") operate in conjunction with each other, and the sensor data (variables) collected by sensors attached to those facility components also have some correlation. Therefore, if an abnormality occurs in one of the multiple facility components that make up the facility and the abnormality is detected, the effect of the abnormality will propagate and the abnormality will be detected in multiple sensor data. In such cases, it is not easy to identify (precisely diagnose) the facility component that is the source of the abnormality. Therefore, a technology has been known in the past that, when multiple equipment components in a facility enter an abnormal operating state, makes it possible to estimate the equipment component that caused the abnormal operating state by using sensor data collected by sensors installed in multiple equipment components within the facility. For example, Patent Document 1 discloses an abnormality diagnosis system that estimates the part causing the change in the state of equipment based on state changes based on changes in the relationships between multiple operating data related to the target parts detected by multiple detection units set for each part (device) of the equipment, and on detection unit relationship information that stores the propagation relationships of effects between the parts of the equipment corresponding to the detection units. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2017 / 159016 Summary of the Invention [Problem to be solved by the invention]

[0004] In conventional techniques such as that disclosed in Patent Document 1, an operator or the like needs to grasp in advance the propagation relationships of the effects between a plurality of equipment components, and it is necessary to be able to prepare information equivalent to the inter-detection unit relationship information in which the propagation relationships are defined based on the grasped propagation relationships. In contrast, for example, in a complex facility with feedback control or in a large-scale facility, it is difficult for an operator or the like to grasp the propagation relationship of influences between devices that make up the facility. Therefore, with conventional technologies, there was a problem that it was not possible to grasp the propagation relationship of effects between multiple equipment components, or even if it was possible to grasp it, the accuracy was low, making it possible to not estimate the cause of an abnormality that occurred in the equipment.

[0005] The present disclosure has been made to solve the problems described above, and aims to provide an abnormality cause estimation device that is capable of estimating the cause of an abnormality that occurs in equipment, regardless of the complexity or size of the equipment. [Means for solving the problem]

[0006] The anomaly cause estimation device according to the present disclosure includes a sensor data acquisition unit that acquires multiple time-series sensor data collected by multiple sensors provided in multiple equipment components that constitute a target equipment; an anomaly detection unit that detects multiple anomaly detection sensors in which an anomaly has occurred among the multiple sensors based on the multiple sensor data acquired by the sensor data acquisition unit; an anomaly detection order estimation unit that estimates an anomaly detection order in which an anomaly was detected for the multiple anomaly detection sensors based on detection times at which the anomaly detection unit detected the multiple anomaly detection sensors; an anomaly propagation path tracing unit that estimates an anomaly propagation order in which an anomaly has propagated based on anomaly detection sensor information regarding the multiple anomaly detection sensors detected by the anomaly detection unit and an estimated structure in which dependency relationships between the equipment components are indicated; and an anomaly cause estimation unit that estimates a cause of the anomaly based on the anomaly detection order estimated by the anomaly detection order estimation unit and the anomaly propagation order estimated by the anomaly propagation path tracing unit. Effect of the Invention

[0007] According to the present disclosure, as configured as described above, the abnormality cause estimation device is capable of estimating the cause of an abnormality that has occurred in equipment, regardless of the complexity or size of the equipment. [Brief description of the drawings]

[0008] [Figure 1] 1 is a diagram illustrating an example of the configuration of a precision diagnosis system including an abnormality factor estimating device according to a first embodiment. [Diagram 2] 1 is a diagram illustrating an example of a configuration of an abnormality factor estimating device according to a first embodiment. [Diagram 3] FIG. 4 is a diagram for explaining a configuration of sensor data in the first embodiment. [Figure 4] 10 is a diagram for explaining a specific example of an anomaly propagation order estimation process performed by an anomaly propagation path tracing unit in embodiment 1. FIG. [Diagram 5] 13 is another diagram for explaining a specific example of the anomaly propagation order estimation process performed by the anomaly propagation path tracing unit in embodiment 1. FIG. [Figure 6] 13 is another diagram for explaining a specific example of the anomaly propagation order estimation process performed by the anomaly propagation path tracing unit in embodiment 1. FIG. [Figure 7] 4 is a diagram for explaining a specific example of an abnormality factor inference process performed by an abnormality factor inference unit in the first embodiment. FIG. [Figure 8] 10 is a diagram for explaining an example of an abnormality factor estimation result screen displayed on a display device by an abnormality factor estimation result output unit in the first embodiment. FIG. [Figure 9] 13 is a diagram for explaining another example of the abnormality factor estimation result screen displayed on the display device by the abnormality factor estimation result output unit in the first embodiment. FIG. [Figure 10] FIG. 10A is a diagram showing an example of the content of an anomaly detection order estimation result, FIG. 10B is a diagram showing an example of the content of an anomaly propagation order estimation result, and FIG. 10C is a diagram showing an example of the content of an anomaly factor order estimation result. [Figure 11] FIG. 2 is a diagram illustrating an example of the configuration of a learning device according to the first embodiment. [Figure 12] FIG. 2 is a diagram showing a concept of an example of a learning process in which a related structure learning unit learns related structures in the first embodiment. [Figure 13] 4 is a flowchart for explaining an operation of the abnormality factor estimating device according to the first embodiment. [Figure 14] 4 is a flowchart for explaining the operation of the learning device according to the first embodiment. [Figure 15] 15 is a flowchart for explaining details of the process of step ST23 in FIG. 14. [Figure 16] 16A and 16B are diagrams illustrating an example of a hardware configuration of the abnormality factor estimating apparatus 100 according to the first embodiment. [Figure 17] FIG. 2 is a diagram showing an example of the configuration of a precision diagnosis system in which an abnormality factor estimating device and a learning device have a common sensor data acquiring unit and data storage unit in the first embodiment. [Figure 18] This figure shows an example configuration of a precision diagnosis system in embodiment 1, in which a learning device learns an associated structure for each operating state of a target equipment, and an abnormality factor estimation device estimates an abnormality factor based on the associated structure corresponding to the operating state of the target equipment that has been learned by the learning device. [Figure 19] 3 is a diagram illustrating an example of the configuration of an abnormality factor estimating device including a related structure correction unit in the first embodiment. FIG. [Figure 20] FIG. 2 is a diagram for explaining a concept of an example of a process in which a related structure modification unit modifies a related structure in the abnormality factor estimating device including the related structure modification unit in the first embodiment. [Figure 21] FIG. 2 is a diagram illustrating an example of the configuration of an abnormality factor estimating device including a relationship change estimating unit in the first embodiment. [Figure 22] FIG. 13 is a diagram for explaining the concept of an example of a relationship change order estimation process performed by the relationship change estimation unit based on a learned relationship structure and a relationship structure in an abnormal state when the anomaly factor estimating device according to the first embodiment is configured to include a relationship change estimation unit. [Figure 23]FIG. 2 is a diagram showing an example of the configuration of an abnormality factor estimating device in the first embodiment, which is provided with an abnormality factor equipment inference unit and has a configuration for estimating an abnormality factor on an equipment basis. [Figure 24] FIG. 13 is a diagram for explaining a concept of an example of an apparatus-by-apparatus anomaly factor estimation process that estimates the cause of an anomaly on an apparatus-by-apparatus basis, the anomaly factor equipment estimation unit performing the anomaly factor estimation process on an anomaly factor equipment basis based on the equipment-attached sensor information, the anomaly detection order estimation result, and the anomaly propagation order estimation result when the anomaly factor estimation device according to the first embodiment is configured to include an anomaly factor equipment estimation unit. [Diagram 25] 10 is a diagram showing an example of an abnormality factor device inference result screen displayed on a display device by an abnormality factor inference result output unit in the first embodiment. FIG. [Figure 26] FIG. 11 is a diagram showing an example of the configuration of an abnormality factor estimating device in the first embodiment, which is provided with an associated structure graph output unit and is configured to output associated structure graph display information to a display device. [Figure 27] FIG. 11 is a diagram for explaining an example of a graph screen displayed on a display device by outputting associated structure graph display information from an associated structure graph output unit in a case where the abnormality factor estimating device is configured to include an associated structure graph output unit in the first embodiment. [Figure 28] FIG. 13 is a diagram showing an example of the contents of a related structure. [Figure 29] FIG. 11 is a diagram illustrating an example of the contents of abnormality detection sensor information. [Diagram 30] FIG. 13 is a diagram showing an example of the contents of a common factor order estimation result. [Diagram 31] 10 is a flowchart for explaining an example of an operation of the abnormality factor estimating device in a case where the abnormality factor estimating device is configured to include an association structure graph output unit in the first embodiment. [Diagram 32] 4 is a diagram showing an example of a configuration of a learning device including a learning sensor pair generating unit in the first embodiment. FIG. [Diagram 33] FIG. 11 is a diagram for explaining the concept of an example of a method in which a learning sensor pair generation unit generates sensor pair information based on facility design information when the learning device is configured to include a learning sensor pair generation unit in embodiment 1. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] In order to describe the present disclosure in more detail, embodiments of the present disclosure will be described below with reference to the accompanying drawings.

[0010] Embodiment 1 The anomaly factor estimation device according to the first embodiment is used for general equipment, such as a power plant or an FA (Factory Automation) system, in which some kind of anomaly appears in sensor data collected within the equipment. The sensor data is collected by sensors provided in a plurality of elements (hereinafter referred to as "equipment components") that make up the equipment. In the first embodiment, the equipment components are assumed to be, for example, equipment. One or more sensors 300 are provided in one equipment. In the following first embodiment, for convenience, it is assumed as an example that one equipment is provided with one sensor. For example, the anomaly factor estimation device monitors sensor data collected in a facility to be monitored, that is, a facility to be detected for occurrence of an anomaly (hereinafter referred to as "target facility"), and detects a plurality of sensors (hereinafter referred to as "anomaly detection sensors") that detect the occurrence of an anomaly based on the sensor data. As described above, in a facility, the facility components that make up the facility, here devices, operate in conjunction with each other, and the sensor data collected by the sensors attached to the devices also have some correlation. When an anomaly occurs in one of the plurality of facility components and is detected, the influence of the anomaly propagates, and the anomaly is detected in a plurality of sensor data, in other words, a plurality of sensors. Then, when the abnormality factor estimation device detects the multiple abnormality detection sensors, it estimates the cause of the abnormality based on the sensor data collected by the multiple abnormality detection sensors, and presents information about the estimation result to an operator such as an on-site maintenance worker of the target equipment. For example, the abnormality factor estimation device displays information about the estimation result of the abnormality factor on a display device, and presents it to the operator. The abnormality factor estimation device presents information about the estimation result of the abnormality factor to the operator in a form that allows the operator to understand, for example, the sensor installed in the equipment that caused the abnormality, or the order in which the operator should inspect it. In this way, the abnormality factor estimation device can reduce unnecessary inspection work by the operator and reduce the burden on the operator. In addition, the abnormality factor estimation device can estimate the cause of the abnormality using a quantitative index that does not rely on human subjectivity, and present the basis for the estimation.

[0011] FIG. 1 is a diagram showing an example of the configuration of a precision diagnosis system 1000 including an abnormality factor estimating device 100 according to the first embodiment. The precision diagnosis system 1000 includes an abnormality factor estimating device 100, a learning device 200, a sensor 300, and a display device 400. In the first embodiment, the sensor 300 and the display device 400 are assumed to be included in the precision diagnosis system 1000, but this is merely an example. It is not essential that the precision diagnosis system 1000 includes the sensor 300 and the display device 400, and the sensor 300 and the display device 400 may be included in a system outside the precision diagnosis system 1000 and connected to the precision diagnosis system 1000. 1, for simplicity of explanation, only one sensor 300 is illustrated, but there may be a plurality of sensors 300. The abnormality factor estimation device 100 is connected to the plurality of sensors 300. In addition, there may be a plurality of display devices 400.

[0012] The abnormality factor estimating device 100 is connected to a learning device 200, a sensor 300, and a display device 400. The fault cause estimation device 100 estimates the cause of an fault that has occurred in a target facility (not shown). Specifically, the anomaly cause estimation device 100 detects a sensor 300 in which an anomaly has occurred (hereinafter referred to as an "anomaly detection sensor") based on the sensor data acquired from the sensor 300 and the learned association structure generated by the learning device 200, and estimates the cause of the anomaly that has occurred in the target equipment by tracing the propagation path of the anomaly between the anomaly detection sensors. In the first embodiment, the abnormality factor estimation device 100 “estimates the abnormality factor” means estimating, for each sensor 300, an abnormality factor score indicating the degree of likelihood that the sensor is the source of an abnormality and an abnormality factor order based on the abnormality factor score, and generating information on the abnormality factor score and the abnormality factor order. Then, the abnormality factor estimating device 100 causes the display device 400 to display information related to the estimated abnormality factor. The abnormality factor estimating device 100 and related structures will be described in detail later.

[0013] The learning device 200 estimates the association structure by using sensor data collected by a sensor 300 provided in the target facility during normal operation of the target facility. In the first embodiment, the estimation of the association structure performed by the learning device 200 is also referred to as "learning". In other words, the "learned association structure" used by the abnormality factor estimation device 100 when estimating the cause of an abnormality that occurs in the target facility can be said to be an "estimated structure" that is an association structure estimated by the learning device 200. The association structure is information indicating the dependency relationships of multiple equipment components that constitute the target facility. The association structure indicates the dependency relationships of the equipment components by indicating the dependency relationships of sensors provided in the equipment components. The association structure is, for example, information indicating the dependency relationships of multiple equipment components that constitute the target facility in a matrix. The association structure may be, for example, information indicating the dependency relationships of multiple equipment components that constitute the target facility in a JSON (JavaScript (registered trademark) Object Notification) format, which is a dictionary-type description format. In the first embodiment, as an example, the association structure that constitutes the target facility is information indicating the dependency relationships of multiple equipment components in a matrix. Note that the normal operation of the target facility specifically refers to the normal operation of multiple facility components (devices in this case) that make up the target facility. Therefore, the sensor data collected by the sensor 300 during the normal operation of the target facility specifically refers to the sensor data collected by the sensor 300 provided in each of the multiple devices that make up the target facility when the devices are operating normally. The learning device 200 will be described in detail later.

[0014] The sensors 300 are provided in a plurality of facility components (here, devices) that make up the target facility. The sensor 300 outputs sensor data to the abnormality cause estimating device 100 . The sensor data is, for example, time series data of sensor measurement values ​​for a predetermined time period obtained at predetermined intervals by the sensor 300 provided in each device that is an equipment component constituting the target equipment. The sensor data indicates, for example, at least one sensor measurement value of the opening degree, deviation, rotation speed, conductivity, flow rate, pressure, temperature, concentration, or water level. Note that this is merely an example, and the sensor data may include, for example, a control value such as a command value or a reference value for a predetermined time period obtained at predetermined intervals by a plurality of sensors 300. In the following embodiment 1, the sensor data is assumed to be time series data of at least one of sensor measurement values ​​such as opening, deviation, rotation speed, conductivity, flow rate, pressure, temperature, concentration, or water level, obtained at predetermined intervals by multiple sensors 300 for a predetermined period of time.

[0015] The display device 400 is, for example, a display of a PC (Personal Computer) installed in a management room where an operator performs his / her work. The display device 400 may be, for example, a touch panel display of a tablet terminal carried by an operator.

[0016] First, an abnormality factor estimating device 100 according to the first embodiment will be described. FIG. 2 is a diagram illustrating an example of a configuration of the abnormality factor estimating device 100 according to the first embodiment. In FIG. 2, the learning device 200 is omitted. 2, for simplicity of explanation, only one sensor 300 is illustrated, as in Fig. 1, but this is merely an example. A plurality of sensors 300 may be connected to the abnormality factor estimation device 100. In the first embodiment, it is assumed that a plurality of sensors 300 are connected to the abnormality factor estimation device 100. In the following first embodiment, the plurality of sensors 300 are also simply referred to as sensors 300.

[0017] The abnormality factor estimation device 100 includes a sensor data acquisition unit 10, a data storage unit 20, an abnormality detection unit 30, an abnormality detection order estimation unit 40, an abnormality propagation path tracing unit 50, an abnormality factor estimation unit 60, and an abnormality factor estimation result output unit 70.

[0018] The sensor data acquiring unit 10 acquires sensor data from the sensor 300 . As described above, in embodiment 1, the sensor data is time series data of sensor measurement values ​​(e.g., at least one of the sensor measurement values ​​of opening, deviation, rotation speed, conductivity, flow rate, pressure, temperature, concentration, or water level) for a predetermined period of time obtained at predetermined intervals from sensors 300 provided in multiple devices that are equipment components. For example, if the quantity of sensors 300 is represented by n, the sensors 300 are represented as X1, X2, X3, X4, ..., Xn. Also, if sensor data is collected at time 1, time 2, ..., time t, the sensor data is represented as a two-dimensional data frame with the number of times t as rows and the number of sensors n as columns. The sensor data of the first sensor X1 at time 1 is represented as X11, the sensor data of the second sensor X2 at time 1 as X21, and the sensor data of the first sensor X1 at time 2 as X12 (see FIG. 3). In the following first embodiment, the sensor data acquired by the sensor data acquisition unit 10 is referred to as "sensor data D1". The sensor data acquiring unit 10 causes the data storage unit 20 to store the acquired sensor data D1.

[0019] The abnormality detection unit 30 performs an abnormality detection process on the sensor data D1 that the sensor data acquisition unit 10 has stored in the data storage unit 20. Specifically, the anomaly detection unit 30 performs an anomaly detection process using a known univariate anomaly detection method on the sensor data D1, which is time-series data stored in the data storage unit 20 by the sensor data acquisition unit 10, and detects a plurality of anomaly detection sensors among the sensors 300. In the first embodiment, an anomaly occurring in the sensor 300 means that the value of the sensor data collected by the sensor 300 is abnormal. In other words, the anomaly detection sensor is the sensor 300 in which the value of the sensor data collected by that sensor 300 is abnormal. Note that in the first embodiment, an anomaly occurring in the sensor 300 means that an anomaly has occurred in the device in which the sensor 300 is provided. Known univariate anomaly detection methods include Discord (Non-Patent Document: KEOGH, Eamonn; LIN, Jessica; FU, Ada. Hot sax: Efficiently finding the most unusual time series subsequence. In: Data mining, fifth IEEE international conference on. IEEE, 2005.) and Hotelling's T^2 theory. In the following first embodiment, the plurality of anomaly detection sensors detected by the anomaly detection unit 30 are also simply referred to as "anomaly detection sensors".

[0020] The anomaly detection unit 30 stores in the data storage unit 20 information related to the anomaly detection sensor (hereinafter referred to as “anomaly detection sensor information”) and information related to the detection time when the anomaly detection sensor detected an anomaly (hereinafter referred to as “anomaly detection time information”). In the following first embodiment, the anomaly detection sensor information is referred to as "anomaly detection sensor information D3" and the anomaly detection time information is referred to as "anomaly detection time information D4." The anomaly detection sensor information D3 is information indicating the anomaly detection sensor. The information indicating the anomaly detection sensor is information capable of identifying the anomaly detection sensor, such as an ID assigned to each sensor 300. The anomaly detection time information D4 is information in which information capable of identifying the anomaly detection sensor is associated with the time at which the anomaly detection sensor detected. The anomaly detection unit 30 may combine the anomaly detection sensor information D3 and the anomaly detection time information D4 into information in which information indicating the anomaly detection sensor is associated with the time at which the anomaly detection sensor detected the anomaly (hereinafter referred to as the "anomaly detection result"). In this case, the anomaly detection unit 30 stores the anomaly detection result in the data storage unit 20.

[0021] In addition, here, the abnormality detection unit 30 acquires the sensor data D1 from the sensor data acquisition unit 10 via the data storage unit 20, but this is merely an example. The abnormality detection unit 30 may acquire the sensor data D1 directly from the sensor data acquisition unit 10.

[0022] The anomaly detection order estimation unit 40 acquires the anomaly detection sensor information D3 and the anomaly detection time information D4 stored in the data storage unit 20 by the anomaly detection unit 30, and performs an anomaly detection order estimation process to estimate the order in which an anomaly was detected to have occurred in the anomaly detection sensor, more specifically, the order in which an anomaly was detected to have occurred in the sensor data D1 collected by the anomaly detection sensor (hereinafter referred to as the "anomaly detection order"). In detail, the anomaly detection order estimation unit 40 assigns anomaly detection orders to the anomaly detection sensors in the order of earliest anomaly detection time, based on the anomaly detection sensor information D3 and the anomaly detection time information D4.

[0023] Specifically, the anomaly detection order estimation unit 40 assigns the anomaly detection order on to the sensor Xn detected as an anomaly detection sensor. Here, the anomaly detection order on is a real value. For example, the anomaly detection order estimation unit 40 assigns the anomaly detection order on in ascending order, starting from the sensor Xn with the earliest associated anomaly detection time. For example, when there are multiple sensors Xn with the same associated anomaly detection time, the anomaly detection order estimation unit 40 assigns the same anomaly detection order on to multiple sensors Xn with the same anomaly detection time. For example, the anomaly detection order estimation unit 40 may assign the anomaly detection order "0" to the sensor Xn with the earliest associated anomaly detection time, and thereafter, to the other sensors Xn, assign the elapsed time from the time corresponding to the anomaly detection order "0" as the anomaly detection order on.

[0024] The anomaly detection order estimation unit 40 stores the anomaly detection order assignment result (hereinafter referred to as the "anomaly detection order estimation result") in the data storage unit 20. In the following first embodiment, the anomaly detection order estimation result is referred to as the "anomaly detection order estimation result D5." The anomaly detection order estimation result D5 is information in which information indicating the anomaly detection sensor, the anomaly detection time, and information indicating the anomaly detection order are associated with each other.

[0025] Note that, here, the anomaly detection order estimation unit 40 acquires the anomaly detection sensor information D3 and the anomaly detection time information D4 from the anomaly detection unit 30 via the data storage unit 20, but this is merely an example. The anomaly detection order estimation unit 40 may acquire the anomaly detection sensor information D3 and the anomaly detection time information D4 directly from the anomaly detection unit 30.

[0026] The anomaly propagation path tracing unit 50 acquires the anomaly detection sensor information D3 and the associated structure stored by the anomaly detection unit 30 from the data storage unit 20, and performs an anomaly propagation order estimation process to estimate the order in which the anomaly propagated (hereinafter referred to as the "anomaly propagation order") based on the acquired anomaly detection sensor information D3 and the associated structure. In the anomaly propagation order estimation process, the anomaly propagation path tracing unit 50 estimates the anomaly propagation order for the sensor 300. Note that for a certain sensor 300, the anomaly detection order assigned by the anomaly detection order estimation unit 40 and the anomaly propagation order estimated by the anomaly propagation path tracing unit 50 are not necessarily the same order. For example, due to a problem of time resolution, even if the same order is assigned to the anomaly detection order, the order relationship of the anomaly propagation order may be reversed. Also, for example, due to a problem of accuracy of the anomaly detection time, the anomaly detection order may be assigned an order opposite to the anomaly propagation order. Furthermore, for example, even if the same order is assigned to the anomaly detection order, the order relationship of the anomaly propagation order may change due to a problem with the accuracy of the anomaly detection time. Although it is relatively easy to detect the presence or absence of an anomaly, it is difficult to accurately detect the time of its occurrence, which may cause a problem with the accuracy of the anomaly detection time. The association structure is generated based on at least one statistic between multiple sensor data by learning in the learning device 200, and is stored in the data storage unit 20. In the following first embodiment, the association structure is referred to as an "association structure D2." As described above, in the first embodiment, the association structure D2 is information in which the dependency relationships between multiple devices, which are multiple equipment components that make up the target equipment, are represented in a matrix.

[0027] In detail, the anomaly propagation path tracing unit 50 determines the dependency relationship between the anomaly detection sensors based on the anomaly detection sensor information D3 and the related structure D2, and traces the direction of the dependency relationship to trace the anomaly propagation path. Then, the anomaly propagation path tracing unit 50 assigns an anomaly propagation order, for example, "1" (th), "2" (th), "3" (th), ... from the anomaly detection sensor located upstream of the anomaly propagation, which is considered to be the source of the anomaly, toward the anomaly detection sensor located downstream. The anomaly propagation path tracing unit 50 generates an estimation result of the anomaly propagation order (hereinafter referred to as "anomaly propagation order estimation result") and stores it in the data storage unit 20. In the following embodiment 1, the anomaly propagation order estimation result is referred to as "anomaly propagation order estimation result D6".

[0028] Here, the anomaly propagation order estimation process performed by the anomaly propagation path tracing unit 50 will be described with reference to the drawings, giving a specific example. 4, 5, and 6 are diagrams for explaining a specific example of the anomaly propagation order estimation process performed by the anomaly propagation path tracing unit 50 in the first embodiment.

[0029] <Generation of influence propagation relationship matrix> In the anomaly propagation order estimation process, the anomaly propagation path tracing unit 50 first converts the association structure D2 stored in the data storage unit 20 into an influence propagation relationship matrix indicating the presence or absence of a dependency relationship between sensor data in the form of a matrix, based on the association structure D2 stored in the data storage unit 20. In the first embodiment, the influence propagation relationship matrix is ​​referred to as an "influence propagation relationship matrix D9."

[0030] FIG. 4 is a diagram for explaining the concept of an example of the process in which the anomaly propagation path tracing unit 50 converts the associated structure D2 into an effect propagation relationship matrix D9 in the first embodiment. Here, the relational structure D2 is a three-dimensional array with the first dimension being the type m of statistical indicators, the second dimension being the number n of sensors, and the third dimension being the number n of sensors. Note that the statistical indicators are indicators that describe the dependency relationships between sensor data. Details of the statistical indicators will be described later. A(k) denotes the two-dimensional matrix corresponding to the kth statistical index, and a(k)ij denotes the statistic describing the dependency from the ith sensor data to the jth sensor data in the kth statistical index. Note that the ith sensor data is the sensor data collected by the ith sensor Xi, and the jth sensor data is the sensor data collected by the jth sensor Xj.

[0031] The anomaly propagation path tracing unit 50 performs preprocessing on the above-mentioned association structure D2 to select statistics to be used for tracing the anomaly propagation path (hereinafter referred to as "tracing statistics"), and generates a preprocessed association structure. In the following embodiment 1, the preprocessed association structure is referred to as a "preprocessed association structure D8." The preprocessed association structure D8 has the same data structure as the association structure D2. The anomaly propagation path tracing unit 50 may select only statistics having a large dependency as the tracking statistics, for example. In this case, the anomaly propagation path tracing unit 50 sets a threshold (hereinafter referred to as a "statistics selection threshold") for each type of statistical index, and selects the statistics a(k)ij whose absolute value |a(k)ij| is greater than the statistics selection threshold as the tracking statistics. For example, when the absolute value of the statistics is greater than the statistics selection threshold, the anomaly propagation path tracing unit 50 substitutes the statistics a(k)ij, which is an element of the association structure D2, for the element b(k)ij of the preprocessed association structure D8, and when the absolute value of the statistics is less than the statistics selection threshold, the anomaly propagation path tracing unit 50 substitutes "0", which indicates that there is no dependency, for the element b(k)ij of the preprocessed association structure D8. The statistical quantity selection threshold may be set manually in advance by an operator or the like by operating an input device (not shown) such as a mouse or a keyboard, or may be automatically determined based on sensor data by the anomaly propagation path tracing unit 50. For example, the anomaly propagation path tracing unit 50 can determine a relative statistical quantity selection threshold from the average value (median, quantile, etc.) of the statistical quantity.

[0032] The anomaly propagation path tracing unit 50 performs a conversion process for converting the pre-processed relation structure D8 into an effect propagation relation matrix D9, and generates the effect propagation relation matrix D9. The anomaly propagation path tracing unit 50 determines the dependency relationship between the sensor data based on one or more types of statistical indicators including at least one directional statistical indicator, and converts the preprocessed association structure D8 into an influence propagation relationship matrix D9. In this case, as a method of determining the dependency relationship between the sensor data, the anomaly propagation path tracing unit 50 may determine that there is a dependency relationship between the sensor data when at least one type of statistical quantity among the statistics b(1)ij, b(2)ij, . . ., b(m)ij is non-"0", or may determine that there is a dependency relationship between the sensor data when the statistics b(1)ij, b(2)ij, . . ., b(m)ij are all non-"0". In detail, for example, when i=1 and j=2, if at least one type of statistical quantity among the statistics b(1)12, b(2)12, ..., b(m)12 is non-"0", it may be determined that there is a dependency between the sensor data collected by the sensor X1 and the sensor data collected by the sensor X2, or if the statistical indicators b(1)12, b(2)12, ..., b(m)12 are all non-"0", it may be determined that there is a dependency between the sensor data collected by the sensor X1 and the sensor data collected by the sensor X2.

[0033] The statistical index for the anomaly propagation path tracing unit 50 to determine the dependency relationship between sensor data may be manually selected by an operator or the like from among m types of statistical indexes. As a specific example, the anomaly propagation path tracing unit 50 causes the display device 400 to display a setting screen for types of statistical indexes on which check boxes for each type of statistical index are displayed. The operator or the like operates an input device such as a mouse or a keyboard to select a statistical index from the setting screen. The anomaly propagation path tracing unit 50 accepts the statistical index selected by the operator or the like as a statistical index for determining the dependency relationship between sensor data.

[0034] In the first embodiment, the influence propagation relation matrix D9 is, for example, a two-dimensional matrix in which the first dimension is the quantity n of the sensors 300 and the second dimension is the quantity n of the sensors 300, as shown in FIG. Here, the elements of the association structure D2 and the preprocessed association structure D8 are real values, and the elements of the influence propagation relationship matrix D9 are Boolean values. For example, in the association structure D2 and the preprocessed association structure D8, the larger the absolute values ​​of the elements |a(k)ij| and |b(k)ij|, the stronger the dependency relationship is, and in the influence propagation relationship matrix D9, when the element cij is "1", it indicates that there is a dependency relationship. For example, if there is a dependency relationship from the i-th sensor data to the j-th sensor data, the anomaly propagation path tracking unit 50 assigns “1” to cij, which is an element of the impact propagation relationship matrix D9, and if there is no dependency relationship from the i-th sensor data to the j-th sensor data, the anomaly propagation path tracking unit 50 assigns “0” to cij, which is an element of the impact propagation relationship matrix D9.

[0035] Estimation of the order of anomaly propagation In the anomaly propagation order estimation process, the anomaly propagation path tracing unit 50 generates the influence propagation relationship matrix D9, and then estimates the anomaly propagation order based on the anomaly detection sensor information D3 acquired from the data storage unit 20 and the generated influence propagation relationship matrix D9.

[0036] 5 and 6 are diagrams for explaining the concept of an example of a process in which the anomaly propagation path tracing unit 50 estimates the anomaly propagation order based on the anomaly detection sensor information D3 and the effect propagation relationship matrix D9 in the first embodiment. Here, as an example, the quantity of sensors 300 is 6, and the sensors 300 are represented by sensors Xn (n=1, . . . , 6). As another example, among sensors X1, X2, X3, X4, X5, and X6, sensors X1, X2, X4, and X6 are anomaly detection sensors. As another example, the effect propagation relationship matrix D9 is a two-dimensional matrix of 6 rows by 6 rows that represents the dependency relationship between sensor data related to sensor Xn.

[0037] First, the anomaly propagation path tracing unit 50 converts the influence propagation relationship matrix D9 into an influence propagation graph D10. As shown in FIG. 5, the influence propagation graph D10 is a directed graph in which the sensors X1, X2, X3, X4, X5, and X6 are represented as nodes, and the dependencies between the sensor data related to the sensors X1, X2, X3, X4, X5, and X6 are represented as edges. The sensors X1, X2, X3, X4, X5, and X6 correspond to the nodes N51, N52, N53, N54, N55, and N56, respectively. For example, when there is a dependency in one direction from the sensor X2 to the sensor X1, the dependency is represented in the influence propagation graph D10 by a one-sided arrow edge from the node N52 to the node N51.

[0038] After converting the effect propagation relationship matrix D9 into an effect propagation graph D10, the anomaly propagation path tracing unit 50 then converts the effect propagation graph D10 into an anomaly propagation graph D11 based on the anomaly detection sensor information D3. For example, as shown in FIG. 5, the anomaly propagation path tracing unit 50 selects only the dependency relationship related to the anomaly detection sensor Xn from among the dependency relationships represented in the influence propagation graph D10, and converts it into an anomaly propagation graph D11. In this case, the anomaly propagation path tracing unit 50 selects only the nodes corresponding to the anomaly detection sensors X1, X2, X4, X6 from among the nodes corresponding to the sensors X1, X2, X3, X4, X5, and X6 represented in the influence propagation graph D10, and the edges between the nodes corresponding to the sensors X1, X2, X3, X4, X5, and X6. For example, in the example shown in FIG. 5, the sensor X3 is not included in the anomaly detection sensors X1, X2, X4, and X6. Therefore, the anomaly propagation path tracing unit 50 does not select the node N53 corresponding to the sensor X3. As a result, in the fault propagation graph D11, the node N53 corresponding to the sensor X3 is deleted, and the edge between the node N53 and the node N54 linked to the node N53 is also deleted.

[0039] Then, the anomaly propagation route tracing unit 50 estimates the anomaly propagation order based on the anomaly propagation graph D11. For example, the anomaly propagation path tracing unit 50 assigns the anomaly propagation order on to the anomaly detection sensor Xn. Here, the anomaly propagation order on is a real value. The anomaly propagation path tracing unit 50 assigns the anomaly propagation order on in ascending order, for example, from the anomaly detection sensor Xn located upstream of the anomaly propagation, which is considered to be the source of the anomaly.

[0040] In the example shown in FIG. 5, the anomaly propagation path tracing unit 50 first determines a node to which no one-sided arrow is drawn from other nodes as the node located most upstream in the anomaly propagation, and assigns the smallest anomaly propagation order "on" to the anomaly detection sensor Xn corresponding to the node. In FIG. 5, the nodes N52, N54, and N56 are nodes to which no one-sided arrow is drawn from other nodes. Therefore, the anomaly propagation path tracing unit 50 assigns anomaly propagation orders o2, o4, and o6 to the anomaly detection sensors X2, X4, and X6 corresponding to the nodes N52, N54, and N56, respectively. At this time, o2=o4=o6. For example, the anomaly propagation path tracing unit 50 sets o2=o4=o6=1. That is, the anomaly propagation path tracing unit 50 sets the anomaly propagation orders of the anomaly detection sensors X2, X4, and X6 to "1" (number one).

[0041] Next, the anomaly propagation path tracing unit 50 assigns an anomaly propagation order on that is larger than the already assigned anomaly propagation order on to the anomaly detection sensor Xn corresponding to the node at the end of the one-sided arrow that comes out from the node to which the smallest anomaly propagation order on (here, "1") is assigned. Also, the anomaly propagation path tracing unit 50 assigns an anomaly propagation order on to the anomaly detection sensor Xn corresponding to the node at the end of the two-sided arrow that comes out from the node to which the smallest anomaly propagation order on (here, "1") is assigned. In FIG. 5, the node N51 is the node at the end of the one-sided arrow that comes out from the node N52. Therefore, the anomaly propagation path tracing unit 50 assigns an anomaly propagation order o1, that is, an anomaly propagation order o1 that is larger than "1" (the number), to the anomaly detection sensor X1 corresponding to the node N51. For example, the anomaly propagation path tracing unit 50 sets the anomaly propagation order o1=2. That is, the anomaly propagation path tracing unit 50 sets the anomaly propagation order of the anomaly detection sensor X1 to "2" (the number). In FIG. 5, the double-headed arrow extending from node N52 ends at node N56, to which the fault propagation order o6=1 has already been assigned.

[0042] Thereafter, the anomaly propagation path tracing unit 50 repeats the above-described assignment of the anomaly propagation order "on" until the anomaly propagation order "on" is assigned to the anomaly detection sensors Xn corresponding to all nodes on the anomaly propagation graph D11. However, there may be a plurality of abnormality propagation orders on that are assigned depending on the path for tracing the abnormality propagation, that is, the selection of the nodes to be traced. For example, in the example shown in FIG. 5, it is assumed that the double-sided arrow between the node N52 and the node N56 is a single-sided arrow from the node N56 to the node N52. In this case, there are two candidates for the abnormality propagation order o1 to be assigned to the abnormality detection sensor X1 corresponding to the node N51. Specifically, as candidates for the abnormality propagation order o1 to be assigned to the abnormality detection sensor X1, there are an abnormality propagation order o1 assigned based on the path from the node N54 directly to the node N51, and an abnormality propagation order o1 assigned based on the path from the node N56 to the node N51 via the node N52. In this case, the abnormality propagation path tracing unit 50 assigns, for example, the later of the candidates for the abnormality propagation order o1 to the abnormality propagation order o1. This is merely an example, and the anomaly propagation path tracing unit 50 may assign the earlier order to the anomaly propagation order o1 of the node N51, for example.

[0043] The anomaly propagation route tracing unit 50 generates an estimation result of the anomaly propagation order being on (hereinafter referred to as the "anomaly propagation order estimation result") and stores it in the data storage unit 20. In the following embodiment 1, the anomaly propagation order estimation result is referred to as the "anomaly propagation order estimation result D6." The anomaly propagation order estimation result D6 is, for example, information indicating the anomaly detection sensor Xn (shown as D6A in FIG. 5), anomaly detection sensor flag fn (shown as D6B in FIG. 5) indicating whether the anomaly detection sensor Xn is included in the anomaly detection sensors, and anomaly propagation order on (shown as D6C in FIG. 5) estimated by the anomaly propagation path tracing unit 50, which are associated with each other. In the example shown in FIG. 5, the sensor Xn included in the anomaly propagation order estimation result D6 is only the anomaly detection sensor Xn. The anomaly detection sensor flag fn is a Boolean value. In the example shown in FIG. 5, the anomaly detection sensor Xn set in the anomaly propagation order estimation result D6, specifically, the anomaly detection sensors X1, X2, X4, and X6, are all anomaly detection sensors, so the anomaly propagation path tracing unit 50 assigns (True) to the anomaly detection sensor flag fn of the anomaly detection sensors X1, X2, X4, and X6, for example.

[0044] In the estimation of the anomaly propagation order by the anomaly propagation path tracing unit 50 based on the anomaly detection sensor information D3 and the influence propagation relationship matrix D9 described using FIG. 5, when converting the influence propagation graph D10 into the anomaly propagation graph D11, the anomaly propagation path tracing unit 50 selected only the dependency relationships between the sensor data related to the anomaly detection sensor Xn, but this is merely one example. For example, as shown in FIG. 6, when converting the influence propagation graph D10 into the anomaly propagation graph D11, the anomaly propagation path tracking unit 50 may select a dependency relationship between sensor data related to a sensor Xn when at least one of two different sensors Xn is an anomaly detection sensor Xn. In this case, the anomaly propagation path tracing unit 50 selects, among the nodes N51, N52, N53, N54, N55, and N56 corresponding to the sensors X1, X2, X3, X4, X5, and X6 represented in the influence propagation graph D10, and the edges between the nodes N51, N52, N53, N54, N55, and N56 corresponding to the sensors X1, X2, X3, X4, X5, and X6, an edge in which at least one of the connected nodes is a node corresponding to the anomaly detection sensors X1, X2, X4, and X6, and a node connected to the edge. For example, the node N53 corresponding to the sensor X3 and the node N54 corresponding to the sensor X4 are connected by an edge of a double-sided arrow. Here, the sensor X3 is not included in the anomaly detection sensor Xn, but the sensor X4 is included in the anomaly detection sensor Xn. Therefore, the edge between the node N53 and the node N45 is not deleted in the anomaly propagation graph D11.

[0045] In this case, for example, the anomaly propagation path tracing unit 50 may generate the anomaly propagation graph D11 so that the nodes N53 and N55 corresponding to the sensors X3 and X5 that are not included in the anomaly detection sensor Xn among the nodes N51, N52, N53, N54, and N55 corresponding to the sensors X1, X2, X3, X4, X5, and X6 shown in the anomaly propagation graph D11 can be understood. In the anomaly propagation graph D11 shown in Fig. 6, the nodes N51, N52, N54, and N56 are shown by solid-line circles, and the nodes N53 and N55 are shown by dotted-line circles.

[0046] In this case, the anomaly propagation path tracing unit 50 estimates the anomaly propagation order on based on the anomaly propagation graph D11 including nodes N53 and N55 corresponding to sensors X3 and X5 that are not included in the anomaly detection sensor Xn among the sensors Xn. For example, in Fig. 6, nodes N53 and N54 do not have one-sided arrows drawn from other nodes. Therefore, the anomaly propagation path tracing unit 50 assigns anomaly propagation orders o3 and o4 to sensors X3 and X4, respectively. In this case, o3 = o4. For example, the anomaly propagation path tracing unit 50 sets o3 = o4 = 1. That is, the anomaly propagation path tracing unit 50 sets the anomaly propagation orders o3 and o4 of sensors X3 and X4 to "1" (order), respectively. Next, the anomaly propagation path tracing unit 50 assigns anomaly propagation orders o1 and o5, which are greater than the anomaly propagation order o4, to the sensors X1 and X5 corresponding to the nodes N51 and N55 at the end points of the one-sided arrow line coming out of the node N54 to which the smallest anomaly propagation order (here, "1") has been assigned. For example, the anomaly propagation path tracing unit 50 sets o1=o5=2. That is, the anomaly propagation path tracing unit 50 sets the anomaly propagation orders of the sensors X1 and X5 to "2" (number). Furthermore, the anomaly propagation path tracing unit 50 assigns anomaly propagation orders o2 and o6, which are greater than the anomaly propagation order o5, to the sensors X2 and X6 corresponding to N52 and N56 at the end points of the one-sided arrow line coming out of the node N55. For example, the anomaly propagation path tracing unit 50 sets o2=o6=3. That is, the anomaly propagation path tracing unit 50 sets the anomaly propagation orders of the sensors X2 and X6 to "3" (number).

[0047] Although X1 corresponding to node N51 has already been assigned the abnormality propagation order o1, in other words, "2nd" (number), node N51 is the node at the end point of the one-sided arrow coming out from node N52, so the abnormality propagation path tracing unit 50 may finally reassign the abnormality propagation order o2, in other words, a number greater than "3rd" (number), as the abnormality propagation order o1 to the sensor X1 corresponding to node N51. For example, the abnormality propagation path tracing unit 50 may reassign the abnormality propagation order o1 to "4th" (number).

[0048] In the example shown in FIG. 6, the nodes N51, N52, N53, N54, N55, and N56 on the anomaly propagation graph D11 include the nodes N53 and N55 corresponding to the sensors X3 and X5 that are not included in the anomaly detection sensor Xn. Therefore, the anomaly propagation path tracing unit 50 may assign the anomaly propagation orders o1, o2, o3, o4, o5, and o6 to the sensors X1, X2, X3, X4, X5, and X6, respectively, and then weight the anomaly propagation orders o3 and o5 assigned to the sensors X3 and X5. For example, in the example shown in FIG. 6, the weight b is added to the anomaly propagation orders o3 and o5 corresponding to the sensors X3 and X5 included in the anomaly propagation order estimation result D6. The weight b is a real value equal to or greater than 0.

[0049] Then, the anomaly propagation route tracing unit 50 causes the data storage unit 20 to store the anomaly propagation order estimation result D6. 6, the sensors Xn, specifically, the sensors X1, X2, X3, X4, X5, and X6 set in the anomaly propagation order estimation result D6 include sensors X3 and X5 that are not included in the anomaly detection sensor Xn (specifically, the anomaly detection sensors X1, X2, X4, and X6). The anomaly propagation path tracing unit 50 assigns (False) to the anomaly detection sensor flags fn of the sensors X3 and X5.

[0050] In this embodiment, the anomaly propagation path tracing unit 50 acquires the anomaly detection sensor information D3 from the anomaly detection unit 30 via the data storage unit 20, but this is merely an example. The anomaly propagation path tracing unit 50 may acquire the anomaly detection sensor information D3 directly from the anomaly detection unit 30.

[0051] Returning to the description of the configuration example of the abnormality factor estimating device 100 shown in FIG. The anomaly cause estimation unit 60 acquires the anomaly detection order estimation result D5 output by the anomaly detection order estimation unit 40 and the anomaly propagation order estimation result D6 output by the anomaly propagation path tracking unit 50 from the data storage unit 20, and performs an anomaly cause estimation process to estimate the cause of the anomaly based on the anomaly detection order estimated by the anomaly detection order estimation unit 40 and the anomaly propagation order estimated by the anomaly propagation path tracking unit 50. In detail, the anomaly factor estimation unit 60 calculates an anomaly factor score indicating the degree of likelihood of being the source of an anomaly based on the anomaly detection order estimation result D5 and the anomaly propagation order estimation result D6, and assigns an order based on the calculated anomaly factor score (hereinafter referred to as the "anomaly factor order"). The higher the degree of likelihood of being the source of an anomaly, the smaller the anomaly factor score becomes.

[0052] First, the abnormality factor estimation unit 60 acquires the abnormality detection order estimation result D5 and the abnormality propagation order estimation result D6 from the data storage unit 20. The abnormality factor estimation unit 60 calculates a corresponding abnormality factor score for each sensor 300 from the abnormality detection order included in the abnormality detection order estimation result D5 and the abnormality propagation order included in the abnormality propagation order estimation result D6. Here, the abnormality detection order, the abnormality propagation order, and the abnormality factor score are all real values. The abnormality factor estimation unit 60 calculates the corresponding abnormality factor scores for all sensors 300 included in the abnormality detection order estimation result D5 or the abnormality propagation order estimation result D6. The abnormality factor estimation unit 60 calculates a representative value as the abnormality factor score by using, for example, a weighted average of the abnormality detection order and the abnormality propagation order. The abnormality factor estimation unit 60 may also calculate, for example, a representative value such as a minimum value or a maximum value as the abnormality factor score. That is, the abnormality factor estimation unit 60 may calculate, for example, the smaller or larger of the abnormality detection order and the abnormality propagation order as the abnormality factor score. Furthermore, when only one of the anomaly detection order and the anomaly propagation order is set, the anomaly factor estimation unit 60 may weight the anomaly factor score taking into consideration that only the set order is set. For example, when only one of the anomaly detection order and the anomaly propagation order is set, the anomaly factor estimation unit 60 adds a weight b to the anomaly factor score. Note that the weight b is a real value equal to or greater than 0. For example, when the anomaly propagation path tracing unit 50 converts the influence propagation graph D10 into the anomaly propagation graph D11, if the anomaly propagation order is assigned while including a node corresponding to a sensor Xn that is not included in the anomaly detection sensor Xn, a situation may occur in which the anomaly detection order is not set but the anomaly propagation order is set.

[0053] After calculating the abnormality factor score for each sensor 300, the abnormality factor estimation unit 60 then assigns an abnormality factor order based on the calculated abnormality factor score to the sensor 300. Note that the abnormality factor order is a real value. The abnormality factor estimating unit 60 assigns the abnormality factor order to the sensor 300 in ascending order of the abnormality factor order corresponding to the sensor 300, for example, from the sensor 300 with the smallest corresponding abnormality factor score. When there are multiple sensors 300 with the same corresponding abnormality factor score, the abnormality factor estimating unit 60 assigns the same abnormality factor order to the multiple sensors 300, for example.

[0054] When the abnormality factor estimation unit 60 assigns the abnormality factor order to each sensor 300, it generates information on the abnormality factor order assigned to each sensor 300 (hereinafter referred to as the "abnormality factor order estimation result") and stores it in the data storage unit 20. In the following first embodiment, the abnormality factor order estimation result is referred to as the "abnormality factor order estimation result D7." The abnormality factor order estimation result D7 is information that corresponds to information indicating the sensor 300, an abnormality detection sensor flag indicating whether the sensor 300 is an abnormality detection sensor included in the abnormality detection order estimation result D5, an abnormality factor score, and an abnormality factor order. The anomaly detection sensor flag is a Boolean value. For example, if the sensor 300 is an anomaly detection sensor, the anomaly cause estimation unit 60 assigns (True) to the anomaly detection sensor flag corresponding to the sensor 300, and if the sensor 300 is not an anomaly detection sensor, the anomaly cause estimation unit 60 assigns (False) to the anomaly detection sensor flag corresponding to the sensor 300. For example, if information about the sensor 300 is included in the anomaly detection order estimation result D5, the anomaly cause estimation unit 60 may determine that the sensor 300 is an anomaly detection sensor Xi.

[0055] The abnormality factor estimation process performed by the abnormality factor estimation unit 60 as described above will be described below with reference to the drawings, giving a specific example. FIG. 7 is a diagram for explaining a specific example of the abnormality factor inference process performed by the abnormality factor inference unit 60 in the first embodiment. 7, as an example, the number of sensors 300 is n (n=1 to 6), and the n-number of sensors 300 is represented as sensor Xn. Furthermore, among the sensors Xn, sensors X1, X2, X4, and X6 are assumed to be anomaly detection sensors. In the example shown in FIG. 7, the abnormality propagation order estimation result D6 used by the abnormality factor estimation unit 60 in the abnormality factor estimation process is information containing only the abnormality propagation order corresponding to the abnormality detection sensor Xi, as shown in FIG. 5.

[0056] In the example shown in FIG. 7, the abnormality cause estimation unit 60 acquires, from the data storage unit 20, an abnormality detection order estimation result D5 and an abnormality propagation order estimation result D6 for the abnormality detection sensors X1, X2, X4, and X6. Here, in the anomaly detection order estimation result D5, the anomaly detection orders o1, o2, o4, o6 (shown by D5C in FIG. 7) corresponding to the anomaly detection sensors X1, X2, X4, X6 are assumed to be "2nd" (th), "3rd" (th), "1st" (th), and "4th" (th), respectively. Also, in the anomaly propagation order estimation result D6, the anomaly propagation orders o1, o2, o4, o6 (shown by D6C in FIG. 7) corresponding to the anomaly detection sensors X1, X2, X4, X6 are assumed to be "2nd" (th), "1st" (th), "1st" (th), and "1st" (th), respectively.

[0057] Here, the abnormality factor estimation unit 60 calculates the abnormality factor score sn (n=1, 2, 4, 6) corresponding to the abnormality detection sensor Xn by using the average. In this case, the abnormality factor estimation unit 60 calculates the abnormality factor scores s1, s2, s4, s6 corresponding to the abnormality detection sensors X1, X2, X4, X6 as "2", "2", "1", and "2.5", respectively. Based on the calculated abnormality factor scores s1, s2, s4, s6, the abnormality factor estimation unit 60 assigns the abnormality factor orders o1, o2, o4, o6 corresponding to the abnormality detection sensors X1, X2, X4, X6 as "2" (number), "2" (number), "1" (number), and "3" (number), respectively.

[0058] Then, the abnormality factor estimating section 60 generates an abnormality factor order estimation result D7. Specifically, the abnormality factor estimation unit 60 generates an abnormality factor order estimation result D7 in which information indicating the abnormality detection sensors X1, X2, X4, and X6 (shown by D7A in FIG. 7), abnormality detection sensor flags f1, f2, f4, and f6 (shown by D7B in FIG. 7) indicating whether the abnormality detection sensors X1, X2, X4, and X6 are included in the abnormality detection order estimation result D5, abnormality factor scores s1, s2, s4, and s6 (shown by D7C in FIG. 7), and abnormality factor orders o1, o2, o4, and o6 (shown by D7D in FIG. 7) are associated with each other. Here, the sensor Xn included in the anomaly detection order estimation result D5 is equal to the anomaly detection sensor Xn. Therefore, the anomaly detection sensor flags f1, f2, f4, and f6 corresponding to the anomaly detection sensors X1, X2, X4, and X6 are all (True). Then, the abnormality factor estimating section 60 stores the generated abnormality factor order estimation result D7 in the data storage section 20.

[0059] Note that, here, the anomaly factor estimation unit 60 acquires the anomaly detection order estimation result D5 from the anomaly detection order estimation unit 40 via the data storage unit 20, and acquires the anomaly propagation order estimation result D6 from the anomaly propagation path tracing unit 50 via the data storage unit 20, but this is merely an example. The anomaly factor estimation unit 60 may acquire the anomaly detection order estimation result D5 and the anomaly propagation order estimation result D6 directly from the anomaly detection order estimation unit 40 and the anomaly propagation path tracing unit 50, respectively.

[0060] Returning to the description of the configuration example of the abnormality factor estimating device 100 shown in FIG. The abnormality factor estimation result output unit 70 acquires the abnormality factor order estimation result D7 output by the abnormality factor estimation unit 60, the abnormality detection order estimation result D5 output by the abnormality detection order estimation unit 40, and the abnormality propagation order estimation result D6 output by the abnormality propagation path tracking unit 50 from the data storage unit 20, and outputs information related to the estimation results of the abnormality factors by the abnormality factor estimation unit 60. In detail, the abnormality factor estimation result output section 70 outputs information (hereinafter referred to as "abnormality factor estimation result display information") for displaying a screen showing information on the abnormality factor estimation result estimated by the abnormality factor estimation section 60 (hereinafter referred to as "abnormality factor estimation result screen") to the display device 400 based on the abnormality factor order estimation result D7, the abnormality detection order estimation result D5, and the abnormality propagation order estimation result D6. In the first embodiment, the abnormality factor estimation result output unit 70 is provided in the abnormality factor estimation device 100, but this is merely an example. The abnormality factor estimation result output unit 70 may be provided in a device such as a display (not shown) connected to the abnormality factor estimation device 100 via a wired or wireless signal line.

[0061] Here, FIG. 8 and FIG. 9 are diagrams for explaining an example of an abnormality factor estimation result screen displayed on the display device 400 by the abnormality factor estimation result output unit 70 in the first embodiment. 8 and 9 show an example of an abnormality cause estimation result screen in the case where the number of sensors 300 is 6 (sensors Xn, n=1 to 6). 8 and 9, the abnormality cause estimation result screens are indicated by "D12-1" and "D12-2," respectively. For example, the abnormality cause estimation result screen has eleven display frames, as shown in Figures 8 and 9, namely, display frame D12A, display frame D12B, display frame D12C, display frame D12D, display frame D12E, display frame D12F, display frame D12G, display frame D12H, display frame D12I, display frame D12J, and display frame D12K. The abnormality factor estimation result output unit 70 displays, for example, an abnormality factor estimation result list in which information on the estimation results of abnormality factors is listed on an abnormality factor estimation result screen. The abnormality factor estimation result list is, for example, a list in which, for each sensor Xn, information indicating the sensor Xn, information indicating the abnormality detection sensor flag, the abnormality detection time, the abnormality detection order, the abnormality propagation order, the abnormality factor score, and the abnormality factor order are displayed in association with each other. In the abnormality factor estimation result screens shown in Figs. 8 and 9, the abnormality factor estimation result lists are indicated by "D12-1a" and "D12-2a", respectively.

[0062] The example of the anomaly factor inference result screen shown in FIG. 8 is an example of a screen when, for example, the content of anomaly detection order inference result D5 is as shown in FIG. 10A, the content of anomaly propagation order inference result D6 is as shown in FIG. 10B, and the content of anomaly factor order inference result D7 is as shown in FIG. 10C.

[0063] The abnormality factor estimation result output unit 70 displays information indicating the sensor Xn of the abnormality factor order estimation result D7 in a display frame D12A, displays information indicating the abnormality detection sensor flag of the abnormality factor order estimation result D7 in a display frame D12B, displays information indicating the abnormality detection time of the abnormality detection order estimation result D5 in a display frame D12C, displays the abnormality detection order of the abnormality detection order estimation result D5 in a display frame D12D, displays the abnormality propagation order of the abnormality propagation order estimation result D6 in a display frame D12E, displays the abnormality factor score of the abnormality factor order estimation result D7 in a display frame D12F, and The abnormality factor estimation result display information, which displays the abnormality factor order of the normal factor order estimation result D7 in the display frame D12G, displays sort buttons for rearranging the order of the abnormality factor estimation results list in ascending order based on the abnormality detection order of the abnormality detection order estimation result D5, the abnormality propagation order of the abnormality propagation order estimation result D6, and the abnormality factor order of the abnormality factor order estimation result D7 in the display frames D12I, D12J, and D12K, respectively, and displays a check box for receiving an instruction to display only the abnormality detection sensors in the display frame D12H, is output to the display device 400. As a result, the display device 400 displays the abnormality factor estimation result screen as shown in FIG.

[0064] For example, the abnormality factor inference result output unit 70 displays “True” in the display frame D12B if the abnormality detection sensor flag of the abnormality factor order inference result D7 is set to (True), and displays “False” in the display frame D12B if the flag is set to (False). Furthermore, for example, if no value is set in the anomaly detection time or the anomaly detection order of the anomaly detection order estimation result D5, the anomaly cause estimation result output unit 70 displays a blank field.

[0065] For example, in the initial state of the abnormality factor estimation result list, the abnormality factor estimation result output unit 70 displays information indicating the sensor Xn, information indicating the abnormality detection sensor flag, the abnormality detection time, the abnormality detection order, the abnormality propagation order, the abnormality factor score, and the abnormality factor order, which are linked to the sensor Xn, in accordance with the order of the IDs etc. assigned to the sensor Xn. Here, the initial state of the abnormality factor estimation result list refers to the state of the abnormality factor estimation result list when the abnormality factor estimation result output unit 70 causes the display device 400 to display the abnormality factor estimation result list for the first time after, for example, power is turned on.

[0066] The example of the abnormality factor estimation result list shown in FIG. 8 is assumed to be an example of an abnormality factor estimation result list in an initial state. In the initial state of the abnormality factor inference result list, the abnormality factor inference result output unit 70 sets the abnormality factor inference result list to a state in which no instruction has been given to rearrange the data or to display only the abnormality detection sensors, as shown in FIG. 8.

[0067] The operator checks, for example, an anomaly cause estimation result screen as shown in FIG. 8. This allows the operator to grasp information regarding the estimation result of the cause of the anomaly. For example, the operator can identify the device that caused the anomaly from the information of the sensor Xn that caused the anomaly. The operator can also grasp the order in which the device in which the anomaly has occurred should be inspected. This allows the operator to reduce unnecessary inspection work, thereby reducing the burden on the operator.

[0068] Furthermore, when the list of abnormality cause estimation results as shown in FIG. 8 is displayed, the operator can instruct to rearrange the information shown in the list of abnormality cause estimation results. For example, the operator can instruct the rearrangement of the information by operating an input device (not shown) such as a mouse or a keyboard to press the sort buttons of the display frames D12I, D12J, and D12K. For example, when the abnormality factor estimation result output unit 70 receives an instruction to rearrange the information, the abnormality factor estimation result output unit 70 displays the sort button to which the instruction was input, in other words, the pressed sort button, by filling it in black. Then, the abnormality factor estimation result output unit 70 outputs, for example, abnormality factor estimation result display information to the display device 400, which is adapted to display a list of abnormality factor estimation results in which the information to be displayed has been rearranged according to the input instruction. As a result, the list of abnormality factor estimation results displayed on the display device 400 is updated to a list of abnormality factor estimation results in which the displayed information has been rearranged.

[0069] Also, for example, an operator or the like can operate the input device to press a check box displayed in the display frame D12H to input an instruction to display only information related to the anomaly detection sensor in the anomaly factor estimation result list. When an instruction to display only information related to the anomaly detection sensor in the anomaly factor estimation result list is input, the anomaly factor estimation result output unit 70 displays a check in the check box. Then, for example, in accordance with the input instruction, the anomaly factor estimation result output unit 70 outputs, to the display device 400, information for displaying anomaly factor estimation result that displays only information related to the anomaly detection sensor in the anomaly factor estimation result list. As a result, the anomaly factor estimation result list displayed on the display device 400 is updated to an anomaly factor estimation result list that displays only information related to the anomaly detection sensor.

[0070] FIG. 9 is a diagram showing an example of an abnormality factor estimation result screen in which, with the abnormality factor estimation result screen as shown in FIG. 8 being displayed, the operator presses a sort button in display frame D12K of the abnormality factor estimation result list, and the abnormality factor estimation result output unit 70, upon receiving this, rearranges the information in the abnormality factor estimation result list in ascending order of abnormality factor order, and then the operator presses a check box in display frame D12H, and the abnormality factor estimation result output unit 70, upon receiving this, displays a list of abnormality factor estimation results in which only rows corresponding to an abnormality detection sensor whose abnormality detection sensor flag is (True) are displayed.

[0071] The abnormality factor estimation result screen shown in FIG. 9 displays only the rows corresponding to sensors X1, X2, X4, and X6 (i.e., abnormality detection sensors) whose abnormality detection sensor flags are (True), and displays a list of abnormality factor estimation results in which the rows corresponding to sensors X1, X2, X4, and X6 are rearranged in ascending order of abnormality factor order based on the abnormality factor order, such as the row corresponding to sensor X4, the row corresponding to sensor X2, the row corresponding to sensor X6, and the row corresponding to sensor X1. On the abnormality factor estimation result screen shown in Fig. 9, a check is displayed in the check box in the display frame D12H. This allows the operator to recognize that only the abnormality detection sensors are displayed on the abnormality factor estimation result screen. Also, on the abnormality factor estimation result screen shown in Fig. 9, the sort button in the display frame D12K is filled in. This allows the operator to recognize that the rows of the abnormality factor estimation result list have been sorted in ascending order of abnormality factor order on the abnormality factor estimation result screen.

[0072] In addition, when the operator presses the same sort button again in a state where any one of the sort buttons displayed in the display frames D12H, D12J, and D12K is pressed, that is, in a state where the display of the abnormality factor estimation result list is rearranged, the display of the abnormality factor estimation result list can be returned to the state before the rearrangement of the display of the abnormality factor estimation result list was instructed. When the abnormality factor estimation result output unit 70 receives the notice that the same sort button has been pressed again, it displays the sort button that has been displayed by filling it in without filling it in. Then, the abnormality factor estimation result output unit 70 outputs, to the display device 400, abnormality factor estimation result display information for displaying the abnormality factor estimation result list before rearrangement. As a result, the abnormality factor estimation result list displayed on the display device 400 is updated to the abnormality factor estimation result list before rearrangement (see FIG. 8).

[0073] Furthermore, for example, when the check box in the display frame D12H is pressed, that is, when the list of abnormality factor estimation results displaying only information related to the abnormality detection sensor is displayed, the operator can return the display of the list of abnormality factor estimation results to the state before the instruction to display only information related to the abnormality detection sensor was given by pressing the check box again. When the abnormality factor estimation result output unit 70 receives the notice that the check box has been pressed again, it displays the unchecked check box. Then, the abnormality factor estimation result output unit 70 outputs, to the display device 400, information for displaying the list of abnormality factor estimation results before switching to display only information related to the abnormality detection sensor. As a result, the list of abnormality factor estimation results displayed on the display device 400 is updated to the list of abnormality factor estimation results before switching to display only information related to the abnormality detection sensor (see FIG. 8).

[0074] Note that, although the initial abnormality factor estimation result screen is as shown in Fig. 8 here, this is merely an example. For example, on the initial abnormality factor estimation result screen, the abnormality factor estimation result output unit 70 may rearrange information related to the sensors 300 displayed in the abnormality factor estimation result list in ascending order based on any one of the abnormality detection order, the abnormality propagation order, and the abnormality factor order, or may display an abnormality factor estimation result list in which only the abnormality detection sensors are displayed.

[0075] Returning to the description of the configuration example of the abnormality factor estimating device 100 shown in FIG. The data storage unit 20 stores various types of information. In detail, the data storage unit 20 stores, for example, the associated structure D2 generated by the learning device 200, the sensor data D1 acquired by the sensor data acquisition unit 10, the anomaly detection sensor information D3 and the anomaly detection time information D4 output by the anomaly detection unit 30, the anomaly detection order estimation result D5 output by the anomaly detection order estimation unit 40, the anomaly propagation order estimation result D6 output by the anomaly propagation path tracking unit 50, and the anomaly factor order estimation result D7 output by the anomaly factor estimation unit 60. 2, the data storage unit 20 is provided in the abnormality factor estimating device 100, but this is merely an example. The data storage unit 20 may be provided in a location outside the abnormality factor estimating device 100 that can be referenced by the abnormality factor estimating device 100.

[0076] An example of the configuration of the learning device 200 according to the first embodiment will be described. FIG. 11 is a diagram illustrating an example of a configuration of the learning device 200 according to the first embodiment. The learning device 200 learns using sensor data collected by a sensor 300 provided in the target equipment during normal operation of the target equipment. In particular, the learning device 200 estimates the associated structure D2 using the sensor data collected by a sensor 300 provided in the target equipment during normal operation of the target equipment. Note that the normal operation of the target equipment refers, in particular, to the normal operation of multiple devices constituting the target equipment. Therefore, the sensor data collected by the sensor 300 during normal operation of the target equipment refers, in particular, to the sensor data collected by the sensor 300 provided in each device during normal operation of the multiple devices constituting the target equipment. The learning device 200 stores the learned association structure D2 in the data storage unit 20 of the abnormality factor estimating device 100. For ease of explanation, FIG. 11 illustrates only the data storage unit 20 as a component of the abnormality factor estimating device 100.

[0077] The learning device 200 includes a learning sensor data acquisition unit 210 , a learning data storage unit 220 , a learning preprocessing unit 230 , and a related structure learning unit 240 .

[0078] The learning sensor data acquisition unit 210 acquires learning data used for learning the related structure D2. The learning data includes sensor data acquired from a plurality of sensors 300. It should be noted that not all of the learning data acquired by the learning sensor data acquisition unit 210 is used for learning the related structure D2. The learning preprocessing unit 230 described below acquires learning data to be actually used for learning the related structure D2 based on the learning data acquired by the learning sensor data acquisition unit 210. Therefore, more accurately, the learning data acquired by the learning sensor data acquisition unit 210 is learning data candidates. Details of the learning preprocessing unit 230 will be described later. The learning sensor data acquiring unit 210 stores the acquired learning data candidates in the learning data storage unit 220 .

[0079] Here, the learning data candidates are the same type of information as the sensor data D1 acquired by the abnormality factor estimation device 100 from the sensor 300, more specifically, information with the same contents as the sensor data D1 (at least one measurement value or control value among the opening degree, deviation, rotation speed, conductivity, flow rate, pressure, temperature, concentration, or water level), and are sensor data collected by the sensor 300 during normal operation of a plurality of devices that are a plurality of equipment components of the target equipment. In other words, the learning data candidates include sensor measurement values ​​or control values ​​during normal operation of the target equipment. Note that the learning data candidates are the same type of information as the sensor data D1 acquired from the same sensor 300 as the sensor 300 from which the abnormality factor estimation device 100 acquired the sensor data D1. The learning data candidates are prepared in advance by, for example, an administrator or the like, and stored in a location that can be referenced by the learning device 200. For example, the learning data candidates may be stored in the data storage unit 20 of the abnormality factor estimation device 100. In this case, the learning sensor data acquisition unit 210 may acquire the learning data candidates from the data storage unit 20 of the abnormality factor estimation device 100. In the first embodiment, the learning data candidates acquired by the learning sensor data acquiring unit 210 are referred to as "learning data candidates D21." 11, the learning data candidates acquired by the learning sensor data acquiring unit 210 from a location other than the abnormality factor estimation device 100 are illustrated as "learning data candidates D21", and the learning data candidates acquired by the learning sensor data acquiring unit 210 from the abnormality factor estimation device 100 are illustrated as "learning data candidates D22". This is for ease of understanding, and the contents of "learning data candidates D21" and "learning data candidates D22" are the same. Therefore, in the following description, both "learning data candidate D21" and "learning data candidate D22" will be described as "learning data candidate D21".

[0080] The learning preprocessing unit 230 acquires the learning data candidates D21 acquired by the learning sensor data acquiring unit 210 from the learning data storage unit 220, and performs preprocessing on the learning data candidates D21. Note that the learning preprocessing unit 230 acquires the learning data candidates D21 stored at the predetermined time from the learning data storage unit 220 at predetermined intervals. In detail, the learning preprocessing unit 230 performs data conversion or selection on the learning data candidates D21, and obtains the learning data to be actually used in learning the relational structure D2.

[0081] For example, the learning preprocessing unit 230 converts the learning data candidates D21 into a first-order difference series and sets the converted learning data candidates D21 as learning data. In this case, the sensor data included in the learning data candidates D21 is converted into data indicating the amount of change. Also, for example, the learning preprocessing unit 230 may select only sensor data with large variance from among the sensor data included in the learning data candidates D21, and use the selected sensor data as the learning data. In this case, the learning preprocessing unit 230 sets a variance threshold, and selects, as the learning data, sensor data with a variance value larger than the threshold from among the sensor data included in the learning data candidates D21. Note that the variance threshold is set manually, for example, by an operator or the like. The operator, for example, operates an input device such as a mouse or a keyboard to input the variance threshold and set the variance threshold. Also, for example, when the measurement error of the sensor 300 is known, the learning preprocessing unit 230 may set the square of the measurement error as the variance threshold.

[0082] The learning preprocessing unit 230 stores the acquired learning data in the learning data storage unit 220. In the first embodiment, the learning data acquired by the learning preprocessing unit 230 is referred to as "learning data D23."

[0083] In the following embodiment 1, as an example, the preprocessing that the learning preprocessing unit 230 performs on the learning data candidates D21 is assumed to be selection, and the learning preprocessing unit 230 acquires, from among the sensor data included in the learning data candidates D21, sensor data whose variance value is greater than a threshold value as the learning data D23.

[0084] It should be noted that, here, the learning preprocessing unit 230 acquires the learning data candidates D21 from the learning sensor data acquiring unit 210 via the learning data storage unit 220, but this is merely an example. The learning preprocessing unit 230 may acquire the learning data candidates D21 directly from the learning sensor data acquiring unit 210.

[0085] The associated structure learning unit 240 acquires the learning data D23 output by the learning preprocessing unit 230 from the learning data storage unit 220, and learns the associated structure D2 based on the learning data D23. In detail, the relational structure learning unit 240 calculates at least one statistic indicating the relationship between two different pieces of sensor data included in the learning data D23, and learns the relational structure D2 based on the calculated statistic.

[0086] The association structure learning unit 240 uses, for example, correlation or cross-correlation, or waveform-based statistical indices such as Granger Causality, Transfer entropy, Convergent cross mapping (CCM), and Dynamic Time Warping (DTW) as indices for calculating statistics indicating the relationship between sensor data (hereinafter referred to as "statistical indices"). The association structure learning unit 240 may also use, for example, Kullback-Leibler (KL) divergence or distribution-based statistical indices such as Histogram Intersection (HI) as statistical indices. Statistical indices are classified into undirected and directed types. An undirected statistical indices refers to a statistical indices that cannot identify the direction of dependency, such as correlation, and a directed statistical indices refers to a statistical indices that can identify the direction of dependency, such as Granger causality. In embodiment 1, as an example, the association structure learning unit 240 calculates one or more types of statistics for multiple sensor data included in the learning data D23, including at least one directional statistical indicator that indicates the relationship between two different sensor data, and learns the association structure D2 based on the calculated statistics.

[0087] The learning process in which the associated structure learning unit 240 learns the associated structure D2 will be described below with a specific example using the drawings. FIG. 12 is a diagram showing a concept of an example of a learning process in which the associated structure learning unit 240 learns the associated structure D2 in the first embodiment.

[0088] As shown in FIG. 12, the learning data D23 is a two-dimensional data frame with rows representing the number of time points t and columns representing the quantity n of sensors 300 (represented by sensors Xn (n=1, 2, 3, 4, . . . , n)). In the association structure D2, all elements defined in the association structure D2 are initialized to "0". Here, it is assumed that the association structure learning unit 240 uses m types of statistical indices when learning the association structure D2.

[0089] <Learning Data Selection> The associated structure learning unit 240 selects sensor data collected by two different sensors Xn from among the sensor data collected by the sensor Xn included in the learning data D23. For example, the associated structure learning unit 240 selects the i-th sensor data collected by the i-th sensor Xi and the j-th sensor data collected by the j-th sensor Xj. Hereinafter, the i-th sensor data collected by the i-th sensor Xi will be simply referred to as the "i-th sensor data" and the j-th sensor data collected by the j-th sensor Xj will be simply referred to as the "j-th sensor data."

[0090] <Statistics calculation> Next, the association structure learning unit 240 calculates the statistics a(1)ij, a(2)ij, . . . , a(m)ij from the i-th sensor data to the j-th sensor data as the statistics from the sensor Xi to the sensor Xj, and calculates the statistics a(1)ij, a(2)ij, . . . , a(m)ij from the i-th sensor data to the j-th sensor data as the statistics from the sensor Xi to the sensor Xj, using m types of statistical indices for the i-th sensor data and the j-th sensor data. Then, the association structure learning unit 240 acquires information on the statistics between the sensor Xi and the sensor Xj (hereinafter referred to as "inter-sensor statistics information") D24A, which is composed of the statistics a(k)ij from the sensor Xi to the sensor Xj and the statistics a(k)ji from the sensor Xj to the sensor Xi. Here, the association structure learning unit 240 converts the inter-sensor statistics information D24A as necessary so that the larger the absolute value |a(k)ij| of the statistics is, the larger the dependency between the sensor data is. For example, the p value corresponding to the Granger Causality statistic a(k)ij takes a value between 0 and 1, and the smaller the p value, the more difficult it is to say that there is no dependency between the i-th sensor data and the j-th sensor data. In this case, the association structure learning unit 240 converts the p value so that the larger the statistic, the greater the dependency, and sets the converted p value (1-p value) as the statistic a(k)ij in the inter-sensor statistics information. On the other hand, if the association structure learning unit 240 calculates the statistics using correlation, which is a non-directional statistical index, for example, the correlation coefficient ρ corresponding to the correlation statistic a(k)ij takes a value between -1 and 1, and the larger the absolute value of ρ, the greater the dependency between the i-th sensor data and the j-th sensor data. In this case, the association structure learning unit 240 does not convert the inter-sensor statistics information D24A.

[0091] In addition, the relational structure learning unit 240 creates pairs of two different sensor data for all combinations of the sensor data included in the learning data D23, calculates statistics for all two different sensor data, and obtains inter-sensor statistical information D24A.

[0092] (Relational Structure Learning) The association structure learning unit 240 learns the association structure D2 using as elements the statistics corresponding to all pairs of sensor data set in the inter-sensor statistics information D24A. For example, the statistic |a(k)ij| from the i-th sensor Xi to the j-th sensor Xj calculated using the k-th statistical index among m types of statistical indexes is substituted for the k-th element in the first dimension, the i-th element in the second dimension, and the j-th element in the third dimension of the association structure D2.

[0093] After learning the relational structure D2 in the above manner, the relational structure learning unit 240 stores the learned relational structure D2 in the data storage unit 20 of the abnormality factor estimating device 100. For example, the associated structure learning unit 240 may store the associated structure D2 in the learning data storage unit 220. In this case, in the anomaly factor estimation device 100, the anomaly propagation path tracing unit 50 downloads the associated structure D2 to be used from the learning data storage unit 220 to the data storage unit 20, for example, every time the anomaly propagation order estimation process is executed.

[0094] In this embodiment, the related structure learning unit 240 acquires the learning data D23 from the learning preprocessing unit 230 via the learning data storage unit 220, but this is merely an example. The related structure learning unit 240 may acquire the learning data D23 directly from the learning preprocessing unit 230.

[0095] Returning to the explanation of the configuration example of the learning device 200 shown in FIG. The learning data storage unit 220 stores various information related to the learning performed by the learning device 200. In detail, the learning data storage unit 220 stores, for example, the learning data candidates D21 acquired by the learning sensor data acquisition unit 210 and the learning data D23 output by the learning preprocessing unit 230. The learning data storage unit 220 may store the associated structure D2 learned by the associated structure learning unit 240. Note that, here, the learning data storage unit 220 is provided in the learning device 200, but this is merely an example, and the learning data storage unit 220 may be provided in a location outside the learning device 200 that can be referenced by the learning device 200.

[0096] In the first embodiment, the learning device 200 includes the learning preprocessing unit 230, but this is merely an example, and the learning device 200 does not necessarily include the learning preprocessing unit 230. When the learning device 200 does not include the learning preprocessing unit 230, the associated structure learning unit 240, for example, sets all of the learning data candidates D21 acquired by the learning sensor data acquiring unit 210 as learning data D23 actually used for learning the associated structure D2, and learns the associated structure D2 using the learning data D23 acquired by the learning sensor data acquiring unit 210. That is, in the learning device 200, the associated structure learning unit 240 sets the multiple learning data candidates acquired by the learning sensor data acquiring unit 210 as multiple learning data, calculates at least one statistic between the multiple learning data based on the multiple learning data, and learns an estimated structure (associated structure D2) in which the dependency between the equipment components is indicated based on the calculated statistic.

[0097] The operations of the abnormality factor estimating device 100 and the learning device 200 according to the first embodiment will be described. First, the operation of the abnormality factor estimating device 100 according to the first embodiment will be described. FIG. 13 is a flowchart for explaining the operation of the abnormality factor estimating apparatus 100 according to the first embodiment.

[0098] The sensor data acquiring unit 10 acquires the sensor data D1 from the sensor 300 (step ST1). The sensor data acquiring unit 10 causes the data storage unit 20 to store the acquired sensor data D1.

[0099] The abnormality detection unit 30 performs an abnormality detection process on the sensor data D1 stored in the data storage unit 20 by the sensor data acquisition unit 10 in step ST1 (step ST2). The abnormality detection unit 30 causes the data storage unit 20 to store the abnormality detection sensor information D3 and the abnormality detection time information D4.

[0100] If the anomaly detection unit 30 detects the anomaly detection sensor in step ST2, the operation of the anomaly factor estimation device 100 proceeds to step ST3. If the anomaly detection unit 30 does not detect the anomaly detection sensor in step ST2, the anomaly factor estimation device 100 ends the process shown in the flowchart in FIG. For example, if the anomaly detection unit 30 detects an anomaly detection sensor in step ST2, it notifies the anomaly detection order estimation unit 40 to that effect, and the operation of the anomaly factor estimation device 100 proceeds to step ST3. On the other hand, if the anomaly detection unit 30 does not detect an anomaly detection sensor in step ST2, it notifies a control unit (not shown) of the anomaly factor estimation device 100 to that effect, and the control unit terminates the processing of the anomaly factor estimation device 100.

[0101] In step ST3, the anomaly detection order estimation unit 40 acquires the anomaly detection sensor information D3 and the anomaly detection time information D4 stored in the data storage unit 20 by the anomaly detection unit 30 in step ST2, and performs an anomaly detection order estimation process to estimate the anomaly detection order in which an anomaly was detected to have occurred in the anomaly detection sensor, more specifically, in which an anomaly was detected to have occurred in the sensor data D1 collected by the anomaly detection sensor (step ST3). The anomaly detection order estimation unit 40 causes the data storage unit 20 to store the anomaly detection order estimation result D5.

[0102] The anomaly propagation path tracking unit 50 acquires the anomaly detection sensor information D3 and associated structure D2 stored by the anomaly detection unit 30 in step ST3 from the data storage unit 20, and performs an anomaly propagation order estimation process to estimate the anomaly propagation order based on the acquired anomaly detection sensor information D3 and associated structure D2 (step ST4). The anomaly propagation path tracing unit 50 outputs the anomaly propagation order estimation result D6 to the data storage unit 20.

[0103] The anomaly cause estimation unit 60 acquires, from the data storage unit 20, the anomaly detection order estimation result D5 output by the anomaly detection order estimation unit 40 in step ST3 and the anomaly propagation order estimation result D6 output by the anomaly propagation path tracking unit 50 in step ST4, and performs an anomaly cause estimation process to estimate the cause of the anomaly based on the anomaly detection order estimated by the anomaly detection order estimation unit 40 and the anomaly propagation order estimated by the anomaly propagation path tracking unit 50 (step ST5). The abnormality factor estimating section 60 generates an abnormality factor order estimation result D7, and stores the generated abnormality factor order estimation result D7 in the data storage section 20.

[0104] The abnormality factor estimation result output unit 70 acquires, from the data storage unit 20, the abnormality factor order estimation result D7 output by the abnormality factor estimation unit 60 in step ST5, the abnormality detection order estimation result D5 output by the abnormality detection order estimation unit 40 in step ST3, and the abnormality propagation order estimation result D6 output by the abnormality propagation path tracking unit 50 in step ST4, and outputs information related to the estimation results of the abnormality factors by the abnormality factor estimation unit 60 (step ST6). In detail, the abnormality factor estimation result output unit 70 outputs abnormality factor estimation result display information for displaying an abnormality factor estimation result screen to the display device 400. In this way, the abnormality factor estimation result output unit 70 presents information on the estimation result of the abnormality factor to the operator.

[0105] If the abnormality factor estimating device 100 does not include the abnormality factor estimation result output unit 70, the abnormality factor estimating device 100 can omit the processing of step ST6 in the operation of the abnormality factor estimating device 100 shown in the flowchart of Fig. 13. The processing of step ST6 is performed, for example, by a device external to the abnormality factor estimating device 100.

[0106] In this way, the anomaly factor estimation device 100 detects multiple anomaly detection sensors based on multiple time-series sensor data collected by multiple sensors 300 installed in the target facility, and estimates the anomaly detection order in which the multiple anomaly detection sensors detected an anomaly based on the detection times at which the multiple anomaly detection sensors detected an anomaly. The anomaly factor estimation device 100 estimates the anomaly propagation order in which the anomaly propagated based on anomaly detection sensor information about the multiple anomaly detection sensors and an estimated structure (association structure) indicating the dependency relationships between multiple component devices that make up the target facility, and estimates the cause of the anomaly based on the estimated anomaly detection order and anomaly propagation order.

[0107] The fault cause estimation device 100 estimates the cause of an abnormality that has occurred in a target facility based on the order in which the abnormalities occurred and the order in which the abnormalities propagate, and therefore can more appropriately estimate the cause of the abnormality based on multiple criteria. In addition, since the abnormality factor estimation device 100 estimates the propagation order of an abnormality based on the estimated structure (association structure), it is possible to estimate the cause of the abnormality early without waiting until sufficient sensor data is collected to construct the association structure D2 when diagnosing the target equipment. In other words, the fault cause estimation device 100 can estimate the cause of an fault that has occurred in the target facility, regardless of the complexity or scale of the target facility.

[0108] Furthermore, the abnormality cause estimating device 100 outputs the estimation result of the abnormality cause. Therefore, the anomaly cause estimation device 100 improves the interpretability and explainability of the estimation results of the anomaly cause to the operator. The anomaly cause estimation device 100 can reduce unnecessary inspection work by the operator and reduce the burden on the operator. Furthermore, the anomaly cause estimation device 100 can estimate the cause of anomaly using a quantitative index that does not rely on human subjectivity and present the basis for the estimation. The operator can determine the inspection sequence of the equipment with little effort.

[0109] Furthermore, the abnormality factor estimating device 100 detects an abnormality using an anomaly detection sensor by using a univariate type anomaly detection method such as the Hotelling theory or Discord. Therefore, the abnormality factor estimation device 100 can more appropriately detect an abnormality such as an independent change in one piece of sensor data D1. An abnormality such as an independent change in one piece of sensor data D1 is, for example, an abnormality detected in the sensor data D1 collected by a single sensor 300 that is not associated with other sensors 300.

[0110] Furthermore, the abnormality factor estimating device 100 detects an abnormality detection sensor using a multivariate type abnormality detection method such as Graphical Lasso. Therefore, the abnormality factor estimation device 100 can more appropriately detect an abnormality that changes the relationship between the multiple sensor data D1. An abnormality that changes the relationship between the multiple sensor data D1 is, for example, an abnormality that occurs in the sensor data D1 collected by the sensor 300 provided in the upstream device when two facility components, here, devices, are in a control relationship, and also appears in the sensor data D1 collected by the sensor 300 provided in the downstream device. For example, when an abnormality occurs in the valve opening detected in a valve that controls some flow rate, an abnormality also appears in the flow rate measured by a flow meter that measures the flow rate.

[0111] Next, the operation of learning device 200 according to the first embodiment will be described. FIG. 14 is a flowchart for explaining the operation of learning device 200 according to the first embodiment.

[0112] The learning sensor data acquiring unit 210 acquires learning data candidates D21 used for learning the related structure D2 (step ST21). Specifically, the learning sensor data acquiring unit 210 acquires the learning data candidates including a plurality of time-series sensor data collected by a plurality of sensors provided in the target equipment during normal operation of the target equipment. The learning sensor data acquiring unit 210 stores the acquired learning data candidates in the learning data storage unit 220 .

[0113] The learning preprocessing unit 230 acquires the learning data candidates D21 acquired by the learning sensor data acquisition unit 210 in step ST21 from the learning data storage unit 220, and performs preprocessing on the learning data candidates D21 to acquire the learning data D23 (step ST22). The learning preprocessing unit 230 stores the acquired learning data D23 in the learning data storage unit 220.

[0114] The related structure learning unit 240 acquires the learning data D23 output by the learning preprocessing unit 230 from the learning data storage unit 220, and learns the related structure D2 based on the learning data D23 (step ST23). When the relational structure D2 is learned, the relational structure learning unit 240 stores the learned relational structure D2 in the data storage unit 20 of the abnormality factor estimating device 100.

[0115] FIG. 15 is a flowchart for explaining the details of the process of step ST23 in FIG.

[0116] Based on the learning data D23 acquired by the learning preprocessing unit 230 in step ST22 of FIG. 14, the associated structure learning unit 240 selects sensor data collected by two different sensors 300 from the sensor data included in the learning data D23. That is, the relational structure learning unit 240 acquires a pair of two different sensor data based on the learning data D23 (step ST231). The relational structure learning unit 240 regards all combinations of a plurality of sensor data included in the learning data D23 as pairs of sensor data.

[0117] The associated structure learning section 240 extracts pairs of sensor data generated in step ST231, and calculates at least one statistic between two different sensor data (step ST232). The associated structure learning unit 240 calculates statistics for all two different sensor data and acquires inter-sensor statistics information D24A.

[0118] Then, the association structure learning unit 240 learns the association structure D2 using the statistics corresponding to all pairs of sensor data set in the inter-sensor statistics information D24A as elements (step ST233).

[0119] If the learning device 200 does not include the learning preprocessing unit 230, the learning device 200 can omit the process of step ST22 in the operation of the learning device 200 shown in the flowchart of FIG.

[0120] In this way, the learning device 200 acquires a plurality of time-series sensor data collected by a plurality of sensors 300 provided in the target facility during normal operation of the target facility as learning data candidates, and acquires a plurality of learning data to be used for learning based on the plurality of learning data candidates. Based on the acquired learning data, the learning device 200 calculates at least one statistic between the plurality of sensor data included in the learning data, and learns an estimated structure (associated structure D2) based on the calculated statistic.

[0121] For example, the comprehensiveness of an association structure taught manually depends on the connection relationships of the equipment components that the person understands, in other words, the sensors 300. In contrast, the learning device 200 is capable of comprehensively extracting the associations between the sensor data D1, and as a result, can provide an association structure D2 that suppresses overlooking of the connection relationships of the sensors 300. The learning device 200 provides the anomaly factor estimation device 100 with an estimation structure (association structure) for tracking the sensor 300 that is the source of the anomaly, thereby allowing the anomaly factor estimation device 100 to more appropriately track the sensor 300 that is the source of the anomaly, thereby improving the estimation accuracy of the cause of the anomaly. Furthermore, even if an association structure that contains only qualitative information such as connection information is manually taught, the learning device 200 can quantitatively determine the magnitude of association or the direction of influence between the sensor data D1, and can generate and provide an estimation structure (association structure) that enables more appropriate estimation of the cause of anomaly.

[0122] Furthermore, the learning device 200 calculates statistics using waveform-based statistical indices such as correlation, Granger Causality, or DTW. Therefore, the learning device 200 can track anomaly propagation based on the dependency of waveform similarity, and can provide an estimation structure (association structure D2) that can more appropriately estimate the cause of the anomaly.

[0123] Furthermore, the learning device 200 calculates statistics using a distribution-based statistical index such as KL divergence or HI. Therefore, the learning device 200 can track anomaly propagation based on a dependency relationship that is distributionally similar, and can provide an estimation structure (association structure D2) that can more appropriately estimate anomaly causes.

[0124] Furthermore, the learning device 200 calculates statistics using waveform-based statistical indices and distribution-based statistical indices. Therefore, the learning device 200 can track anomaly propagation based on a dependency relationship such as waveform or distribution similarity, and can provide an estimation structure (association structure D2) that can more appropriately estimate the cause of the anomaly.

[0125] 16A and 16B are diagrams illustrating an example of a hardware configuration of the abnormality factor estimating apparatus 100 according to the first embodiment. In the first embodiment, the functions of the sensor data acquiring unit 10, the anomaly detecting unit 30, the anomaly detection order estimating unit 40, the anomaly propagation path tracing unit 50, the anomaly factor estimating unit 60, the anomaly factor estimation result output unit 70, and a control unit (not shown) are realized by the processing circuit 1601. That is, the anomaly factor estimating device 100 includes the processing circuit 1601 for estimating the factor of an anomaly that has occurred in the target facility by using an estimation structure (association structure D2) indicating the dependency relationships between a plurality of facility components that constitute the target facility. The processing circuit 1601 may be dedicated hardware as shown in FIG. 16A, or may be a processor 1604 that executes a program stored in a memory as shown in FIG. 16B.

[0126] When the processing circuit 1601 is dedicated hardware, the processing circuit 1601 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0127] When the processing circuit is the processor 1604, the functions of the sensor data acquisition unit 10, the anomaly detection unit 30, the anomaly detection order estimation unit 40, the anomaly propagation path tracing unit 50, the anomaly factor estimation unit 60, the anomaly factor estimation result output unit 70, and the control unit (not shown) are realized by software, firmware, or a combination of software and firmware. The software or firmware is described as a program and stored in the memory 1605. The processor 1604 reads out and executes the program stored in the memory 1605 to execute the functions of the sensor data acquisition unit 10, the anomaly detection unit 30, the anomaly detection order estimation unit 40, the anomaly propagation path tracing unit 50, the anomaly factor estimation unit 60, the anomaly factor estimation result output unit 70, and the control unit (not shown). That is, the anomaly factor estimation device 100 includes a memory 1605 for storing a program that, when executed by the processor 1604, results in the execution of steps ST1 to ST6 in FIG. 13 described above. The program stored in the memory 1605 can be said to cause the computer to execute the procedures or methods of the sensor data acquisition unit 10, the anomaly detection unit 30, the anomaly detection order estimation unit 40, the anomaly propagation path tracking unit 50, the anomaly factor estimation unit 60, the anomaly factor estimation result output unit 70, and the control unit (not shown). Here, the memory 1605 corresponds to, for example, a non-volatile or volatile semiconductor memory such as a RAM, a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable Read Only Memory), an EEPROM (registered trademark; description will be omitted below) (Electrically Erasable Programmable Read-Only Memory), a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, a DVD (Digital Versatile Disc), and the like.

[0128] The functions of the sensor data acquiring unit 10, the anomaly detecting unit 30, the anomaly detection order estimating unit 40, the anomaly propagation path tracing unit 50, the anomaly factor estimating unit 60, the anomaly factor estimation result output unit 70, and a control unit (not shown) may be partially realized by dedicated hardware and partially realized by software or firmware. For example, the functions of the sensor data acquiring unit 10 and the anomaly factor estimation result output unit 70 may be realized by a processing circuit 1601 as dedicated hardware, and the functions of the anomaly detection order estimating unit 40, the anomaly propagation path tracing unit 50, the anomaly factor estimating unit 60, and a control unit (not shown) may be realized by a processor 1604 reading and executing a program stored in a memory 1605. The data storage unit 20 is configured with an auxiliary storage device (not shown). The abnormality factor estimating device 100 also includes an input interface device 1602 and an output interface device 1603 that perform wired or wireless communication with devices such as the sensor 300 or the display device 400.

[0129] An example of the hardware configuration of the learning device 200 according to the first embodiment is also as shown in FIG. 16A and FIG. 16B. In the first embodiment, the functions of the learning sensor data acquisition unit 210, the learning preprocessing unit 230, and the associated structure learning unit 240 are realized by the processing circuit 1601. That is, the learning device 200 includes the processing circuit 1601 for controlling learning of an estimated structure (associated structure D2) indicating dependencies between multiple equipment components constituting the target equipment, based on multiple time-series sensor data collected by multiple sensors 300 provided in the target equipment during normal operation of the target equipment. The processing circuit 1601 may be dedicated hardware as shown in FIG. 16A, or may be a processor 1604 that executes a program stored in a memory as shown in FIG. 16B.

[0130] When the processing circuit 1601 is dedicated hardware, the processing circuit 1601 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0131] When the processing circuit is the processor 1604, the functions of the learning sensor data acquisition unit 210, the learning preprocessing unit 230, and the associated structure learning unit 240 are realized by software, firmware, or a combination of software and firmware. The software or firmware is written as a program and stored in the memory 1605. The processor 1604 executes the functions of the learning sensor data acquisition unit 210, the learning preprocessing unit 230, and the associated structure learning unit 240 by reading and executing the program stored in the memory 1605. That is, the learning device 200 includes the memory 1605 for storing a program that, when executed by the processor 1604, results in the execution of steps ST21 to ST23 in FIG. 14 described above. Also, it can be said that the program stored in the memory 1605 causes a computer to execute the procedure or method of the processing of the learning sensor data acquisition unit 210, the learning preprocessing unit 230, and the associated structure learning unit 240. Here, memory 1605 refers to, for example, non-volatile or volatile semiconductor memory such as RAM, ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), magnetic disk, flexible disk, optical disk, compact disk, mini disk, DVD (Digital Versatile Disc), etc.

[0132] Note that the functions of the learning sensor data acquisition unit 210, the learning preprocessing unit 230, and the related structure learning unit 240 may be partially realized by dedicated hardware and partially realized by software or firmware. For example, the functions of the learning sensor data acquisition unit 210 may be realized by a processing circuit 1601 as dedicated hardware, and the functions of the learning preprocessing unit 230 and the related structure learning unit 240 may be realized by a processor 1604 reading and executing a program stored in a memory 1605. The learning data storage unit 220 is configured with an auxiliary storage device (not shown). The learning device 200 also includes an input interface device 1602 and an output interface device 1603 that perform wired or wireless communication with devices such as the abnormality factor estimating device 100.

[0133] In the above-described first embodiment, it is assumed for convenience that one device is provided with one sensor 300, but this is merely an example, and one device may be provided with multiple sensors 300. For example, in the above-described first embodiment, the equipment components may be multiple parts of the device, and each part may be provided with a sensor 300. In this case, the abnormality factor estimation device 100 presents information on the estimation result of the abnormality factor to the operator in a form that allows the operator to understand, for example, the sensor provided in the part causing the abnormality, or the order in which the operator should inspect the parts.

[0134] <Modification> In the above-described first embodiment, the abnormality factor estimation device 100 and the learning device 200 each include the sensor data acquisition unit 10 and the learning sensor data acquisition unit 210, but this is merely an example. Also, in the above-described first embodiment, the abnormality factor estimation device 100 and the learning device 200 each include the data storage unit 20 and the learning data storage unit 220, but this is merely an example. For example, the abnormality factor estimating device 100 and the learning device 200 may be configured to have a common sensor data acquiring unit and data storage unit and to access each other. FIG. 17 is a diagram showing an example configuration of a precision diagnosis system 1000 in which the abnormality factor estimating device 100 and the learning device 200 are provided with a common sensor data acquiring unit 310 and data storage unit 320 in the first embodiment. 17, the abnormality factor estimation device 100 includes an abnormality detection unit 30, an abnormality detection order estimation unit 40, an abnormality propagation path tracking unit 50, an abnormality factor estimation unit 60, an abnormality factor estimation result output unit 70, and a control unit in addition to the data storage unit 320. However, it is not essential for the abnormality factor estimation device 100 to include the abnormality factor estimation result output unit 70. Also, although not shown in FIG. 17 for simplicity of explanation, the learning device 200 includes a learning preprocessing unit 230 and an associated structure learning unit 240 in addition to the sensor data acquisition unit 310.

[0135] In the configuration example of the precision diagnosis system 1000 shown in FIG. 17 , the abnormality factor estimation device 100 is provided with a data storage unit 320, and the learning device 200 is provided with a sensor data acquisition unit 310, but this is merely an example, and in the precision diagnosis system 1000, the learning device 200 may be provided with the data storage unit 320, and the abnormality factor estimation device 100 may be provided with the sensor data acquisition unit 310. Furthermore, in the precision diagnosis system 1000, either the abnormality factor estimating device 100 or the learning device 200 may be provided with the sensor data acquiring unit 310 and the data storage unit 320.

[0136] <Modification> In the above first embodiment, the abnormality factor estimating device 100 is configured to estimate the factor of an abnormality based on one association structure D2, but this is merely an example. For example, the abnormality factor estimation device 100 may be configured to estimate the abnormality factor based on an associated structure corresponding to the operating state of the target equipment. In this case, the learning device 200 learns the associated structure for each operating state of the target equipment. FIG. 18 is a diagram showing an example configuration of a precision diagnosis system 1000 in embodiment 1 in which the learning device 200 learns an associated structure for each operating state of the target equipment, and the abnormality cause estimation device 100 estimates the cause of an abnormality based on the associated structure corresponding to the operating state of the target equipment that has been learned by the learning device 200. 18, the abnormality factor estimation device 100 includes a sensor data acquisition unit 10, an abnormality detection unit 30, an abnormality detection order estimation unit 40, an abnormality factor estimation unit 60, an abnormality factor estimation result output unit 70, and a control unit, in addition to the data storage unit 20 and the abnormality propagation path tracking unit 50. However, the abnormality factor estimation device 100 does not necessarily include the abnormality factor estimation result output unit 70. Also, although not shown in FIG. 18 for simplicity of explanation, the learning device 200 includes a learning sensor data acquisition unit 210 and a learning preprocessing unit 230 in addition to the learning data storage unit 220 and the associated structure learning unit 240.

[0137] 18, in the learning device 200, the associated structure learning unit 240 acquires equipment operation state information D31 representing the operation state of the target equipment corresponding to the learning data candidate D21. For example, an operator or the like operates an input device such as a mouse or a keyboard to input the equipment operation state information D31, and the associated structure learning unit 240 acquires the equipment operation state information D31 by accepting the input equipment operation state information D31. For example, the associated structure learning unit 240 may acquire a control signal or the like in the target equipment and estimate the operating state from the acquired control signal or the like to acquire the equipment operation state information D31. Then, when the associated structure learning unit 240 learns the associated structure D2 based on the learning data D23 output from the learning preprocessing unit 230, it assigns the acquired equipment operation state information D31 to the learned associated structure D2 to generate an associated structure D32. The associated structure learning unit 240 stores the associated structure D32 to which the equipment operation state information D31 has been assigned in the data storage unit 20 of the abnormality factor estimation device 100. The associated structure learning unit 240 may store the associated structure D32 in the learning data storage unit 220. The learning device 200 performs the above-mentioned learning in accordance with various operating states of the target equipment, and learns the association structure D32 corresponding to the various operating states. In this case, in the operation of the learning device 200 described using the flowchart in Fig. 14, the related structure learning unit 240 acquires the equipment operation state information D31 before the processing of step ST23 is performed, and generates and stores the related structure D32 in step ST23. The learning device 200 repeatedly performs the operation shown in the flowchart in Fig. 14 according to the operating state of the target equipment.

[0138] In the anomaly factor estimation device 100, the anomaly propagation path tracing unit 50 acquires equipment operation state information D31. When performing an anomaly propagation order estimation process, the anomaly propagation path tracing unit 50 estimates the anomaly propagation order based on the acquired anomaly detection sensor information D3, the equipment operation state information D31, and an association structure D32 stored in the data storage unit 20 by the learning device 200. Specifically, the anomaly propagation path tracing unit 50 selects an association structure D32 corresponding to the operation state of the target equipment, and estimates the anomaly propagation order using the selected association structure D32. For example, in the learning device 200, the related structure learning unit 240 may store the related structure D32 in the learning data storage unit 220, and in the anomaly factor estimation device 100, the anomaly propagation path tracking unit 50 may download the related structure D32 to be used from the learning data storage unit 220 to the data storage unit 20 each time the anomaly propagation order estimation process is executed. In this case, in the operation of the anomaly factor estimation device 100 described using the flowchart in Figure 13, in step ST4, the anomaly propagation path tracking unit 50 estimates the anomaly propagation order based on the anomaly detection sensor information D3, the equipment operation state information D31, and the association structure D32 stored in the data storage unit 20 by the learning device 200.

[0139] In this way, the anomaly cause estimation device 100 can be configured so that the anomaly propagation path tracking unit 50 estimates the anomaly propagation order based on the anomaly detection sensor information D3, equipment operation state information D31 indicating the operation state of the target equipment, and an estimation structure (association structure D32) indicating the dependency relationships between multiple target components that make up the target equipment according to the operation state of the target equipment. With such a configuration, the abnormality factor estimation device 100 can respond to changes in the dependency relationships between the sensors 300 that accompany changes in the operating state of the target equipment, and can precisely estimate the causes of abnormalities based on the association structure D32 with improved reliability. In the learning device 200, the related structure learning unit 240 calculates at least one statistical quantity between multiple sensor data based on the learning data, and learns an estimated structure (related structure D2) based on the calculated statistical quantity.By generating a related structure D32 by adding equipment operation status information D31 to the related structure D2, the reliability of the related structure provided to the abnormality factor estimation device 100 is improved, and the related structure D32 that can precisely estimate the cause of an abnormality can be provided to the abnormality factor estimation device 100.

[0140] <Modification> In the above-described first embodiment, the abnormality factor estimating device 100 includes the data storage unit 20, but this is merely an example. For example, one or more network storage devices (not shown) arranged on a communication network may store various data, and in the abnormality factor estimation device 100, the abnormality detection unit 30, the abnormality detection order estimation unit 40, the abnormality propagation path tracking unit 50, the abnormality factor estimation unit 60, and the abnormality factor estimation result output unit 70 may access the network storage device.

[0141] <Modification> In the above-described first embodiment, in the anomaly factor estimation device 100, the anomaly detection unit 30 performs an anomaly detection process on the sensor data D1 using a known univariate anomaly detection method and detects an anomaly detection sensor among the sensors 300, but this is merely one example. For example, the anomaly detection unit 30 may detect the anomaly detection sensor using a known multivariate anomaly detection method. For example, a known multivariate anomaly detection method may be a method such as Graphical Lasso. In addition, the anomaly detection unit 30 may perform an anomaly detection process using, for example, a univariate anomaly detection method and a multivariate anomaly detection method in combination.

[0142] In this way, by detecting anomaly detection sensors using a univariate anomaly detection method, a multivariate anomaly detection method, or both methods, the anomaly cause estimation device 100 can appropriately detect anomaly detection sensors that detect the occurrence of various types of anomalies, even if the way the anomaly is manifested in the sensor data varies depending on the type of anomaly that occurs in the target equipment, and can precisely estimate the cause of the anomaly.

[0143] <Modification> In the above-described first embodiment, the abnormality factor estimating device 100 may be configured to include an associated structure correcting unit 330 that corrects the associated structure D2 stored in the data storage unit 20. FIG. 19 is a diagram showing an example of the configuration of the abnormality factor estimating device 100 including a related structure correction unit 330 in the first embodiment. 19, the abnormality factor estimation device 100 includes a sensor data acquisition unit 10, an abnormality detection unit 30, an abnormality detection order estimation unit 40, an abnormality factor estimation unit 60, an abnormality factor estimation result output unit 70, and a control unit, in addition to the related structure correction unit 330, the data storage unit 20, and the anomaly propagation path tracking unit 50. However, it is not essential that the abnormality factor estimation device 100 includes the abnormality factor estimation result output unit 70. In addition, although not shown in FIG. 19, the abnormality factor estimation device 100 is connected to a learning device 200, a sensor 300, and a display device 400.

[0144] The associated structure correction unit 330 acquires information about sensor pairs having a dependent relationship (hereinafter referred to as "dependent pair information") D33 and information about sensor pairs having no dependent relationship (hereinafter referred to as "non-dependent pair information") D34. The dependent pair information D33 and the non-dependent pair information D34 may be generated manually by an operator based on the operator's know-how, or may be generated by the associated structure correction unit 330 from information indicating a physical connection relationship such as design information of equipment. For example, the dependent pair information D33 and the non-dependent pair information D34 may be generated by combining manual generation based on the operator's know-how and generation by the associated structure correction unit 330 based on information indicating a physical connection relationship such as design information of equipment. The relational structure correction unit 330 acquires the relational structure D2 from the data storage unit 20, corrects the dependency relationship between the sensor data for the relational structure D2 based on the dependency pair information D33 and the dependency free pair information D34, and stores the corrected relational structure D35 in the data storage unit 20. Note that the corrected relational structure D35 has the same data structure as the relational structure D2.

[0145] When the corrected relational structure D35 is stored in the data storage unit 20, the anomaly propagation path tracing unit 50 performs an anomaly propagation order estimation process using the relational structure D35.

[0146] The process of modifying the related structure D2 by the related structure modification unit 330 will be described in detail below with a specific example. FIG. 20 is a diagram for explaining a concept of an example of a process in which the related structure modification unit 330 modifies the related structure D2 in the abnormality factor estimating device 100 including the related structure modification unit 330 in the first embodiment. Here, as an example, the quantity of sensors 300 is 3, and the sensors 300 are represented by sensors Xn (n=1, 2, 3). In addition, the dependent pair information D33 defines that there is a dependent relationship from the i-th sensor Xi to the j-th sensor Xj among the sensors Xn, and the non-dependent pair information D34 defines that there is no dependent relationship from the i-th sensor Xi to the j-th sensor Xj among the sensors Xn. Both i and j are 1, 2, or 3. For example, it is assumed that, among sensors X1, X2, and X3, there is a dependency relationship between the sensor data from sensor X1 to sensor X2 and from sensor X3 to sensor X2, but there is no dependency relationship between the sensor data from sensor X2 to sensor X1. The association structure D2 is a three-dimensional array in which the first dimension is a statistical index, the second dimension is the number of sensors Xn, and the third dimension is the number of sensors Xn. The number of types of statistical indexes is two.

[0147] The related structure correction unit 330 corrects the statistics a(k)ij of the related structure D2 to statistics a'(k)ij based on the dependent pair information D33 and the non-dependent pair information D34. Here, the statistics a(k)ij and the statistics a'(k)ij are real values. For example, the related structure correction unit 330 corrects the statistics a(k)ij of the related structure D2 corresponding to the pair of sensors Xi and Xj included in the dependent pair information D33 to statistics a'(k)ij larger than the upper limit of the statistics, indicating that there is a dependency relationship for each statistical indicator. Note that this is only an example, and for example, the related structure correction unit 330 may correct the statistics a(k)ij of the related structure D2 corresponding to the pair of sensors Xi and Xj included in the dependent pair information D33 to statistics a'(k)ij larger than a threshold value set for selecting a dependency relationship. 20, the association structure correction unit 330 corrects the statistics a(1)12 and a(2)12 corresponding to the sensor data from sensor X1 to sensor X2, which have a dependent relationship, to the statistics a'(1)12 and a'(2)12, which are the upper limit values ​​of the statistical index, respectively, based on the dependency pair information D33. Also, the association structure correction unit 330 corrects the statistics a(1)32 and a(2)32 corresponding to the sensor data from sensor X3 to sensor X2, which have a dependent relationship, to the statistics a'(1)32 and a'(2)32, which are the upper limit values ​​of the statistical index, respectively, based on the dependency pair information D33.

[0148] In addition, the relational structure correction unit 330 corrects the statistics a(k)ij of the relational structure D2 corresponding to the pair of sensors Xi and Xj included in the non-dependence pair information D34 to statistics a'(k)ij indicating that there is no dependency for each statistical indicator. In FIG. 20, the associated structure correction unit 330 corrects the statistics a(1) 21 and a(2) 21 corresponding to the sensor data from sensor X2 to sensor X1, which has no dependency, based on the non-dependence pair information D34, to statistics a'(1) 21 and a'(2) 21, which indicate that there is no dependency, respectively.

[0149] 19, when the abnormality factor estimating device 100 is configured to include the association structure correction unit 330, in the operation of the abnormality factor estimating device 100 described using the flowchart in Fig. 13, the association structure correction unit 330 corrects the dependency relationship between the sensor data for the association structure D2 based on the dependent pair information D33 and the non-dependent pair information D34 before the processing of step ST4 is performed, and stores the corrected association structure D35 in the data storage unit 20. In step ST4, the abnormality propagation path tracing unit 50 estimates the abnormality propagation order based on the abnormality detection sensor information D3, the equipment operation state information D31, and the association structure D35 corrected by the association structure correction unit 330.

[0150] In this way, the abnormality cause estimation device 100 is equipped with an association structure correction unit 330 that corrects the dependency relationship between sensor data for the association structure D2 (estimation structure) based on dependent pair information D33 for a pair of sensors 300 that have a dependent relationship among the sensors 300 and non-dependent pair information D34 for a pair of sensors 300 that do not have a dependent relationship, thereby improving the reliability of the association structure D2 and enabling precise estimation of the cause of the abnormality.

[0151] <Modification> In the above-described first embodiment, the anomaly factor estimating device 100 may be configured to include a relationship change estimation unit 340 that estimates a change in the relationship between sensor data by comparing the learned relationship structure D2 with the relationship structure D36 at the time of the occurrence of an anomaly. The change in the relationship between sensor data is assumed to be, for example, a breakdown in the relationship between the sensor data. The relationship change estimation unit 340 estimates, in the elements of the relationship structure D2, a location where the breakdown in the relationship between the sensor data is large as a location where a change in the relationship between the sensor data has occurred. FIG. 21 is a diagram showing an example of a configuration of the abnormality factor estimating device 100 including a relationship change estimating unit 340 in the first embodiment. 21, the abnormality factor estimation device 100 includes a relationship change estimation unit 340, a data storage unit 20, an abnormality factor estimation unit 60, an abnormality factor estimation result output unit 70, a sensor data acquisition unit 10, an abnormality detection unit 30, an abnormality detection order estimation unit 40, an abnormality propagation path tracking unit 50, and a control unit, in addition to the relationship change estimation unit 340, a data storage unit 20, an abnormality factor estimation unit 60, and an abnormality factor estimation result output unit 70. However, the abnormality factor estimation device 100 does not necessarily include the abnormality factor estimation result output unit 70. Also, although not shown in FIG. 21 for simplicity of explanation, the learning device 200 includes a learning sensor data acquisition unit 210 in addition to the learning data storage unit 220, a learning preprocessing unit 230, and a related structure learning unit 240.

[0152] The relationship change estimation unit 340 acquires the relationship structure D2, the relationship structure D36 in an abnormal state, and the anomaly detection sensor information D3 from the data storage unit 20. Then, the relationship change estimation unit 340 performs a relationship change order estimation process to compare the relationship structure D2 and the relationship structure D36 based on the relationship structure D2, the relationship structure D36, and the anomaly detection sensor information D3, and estimates the order of changes in the relationship between the sensor data (hereinafter referred to as the "relationship change order"). The associated structure D36 at the time of an abnormality is acquired through the following process. In the learning device 200, the learning sensor data acquisition unit 210 acquires sensor data D1 at the time of an abnormality occurrence (including the period during which the abnormality is detected) from the data storage unit 20 of the abnormality factor estimation device 100, and stores the sensor data D1 at the time of an abnormality occurrence in the learning data storage unit 220. The learning preprocessing unit 230 acquires the sensor data D1 at the time of an abnormality occurrence from the learning data storage unit 220, performs preprocessing on the acquired sensor data D1 at the time of an abnormality occurrence, and outputs the preprocessed sensor data D38 at the time of an abnormality occurrence to the associated structure learning unit 240. The learning preprocessing unit 230 may output the preprocessed sensor data D38 at the time of an abnormality occurrence to the associated structure learning unit 240 via the learning data storage unit 220. The associated structure learning unit 240 learns the associated structure D36 at the time of an abnormality based on the preprocessed sensor data D38 at the time of an abnormality occurrence output from the learning preprocessing unit 230. The related structure learning unit 240 may learn the related structure D36 in a similar manner to the learning of the related structure D2. The related structure learning unit 240 learns the related structure D36 at the time of an abnormality using the sensor data D1 at the time of an abnormality as a target. The associated structure learning unit 240 stores the learned associated structure D36 in the abnormal state in the data storage unit 20 of the abnormality factor estimating device 100.

[0153] The relationship change order estimation process performed by the relationship change estimation unit 340 will be described in detail below. First, the relationship change estimation unit 340 acquires the relationship structure D2 and the abnormality-time relationship structure D36 from the data storage unit 20, and acquires a relationship structure (hereinafter referred to as the "relationship change estimation relationship structure") that is a relationship structure corresponding to one type of non-directional statistical index selected to estimate a relationship change from the relationship structure D2 and the abnormality-time relationship structure D36, and a relationship structure in an abnormality (hereinafter referred to as the "abnormality-time relationship structure for relationship change estimation"). Here, the relationship change estimation relationship structure and the abnormality-time relationship structure for relationship change estimation are two-dimensional matrices A(k), A'(k) that extract only the parts corresponding to the k-th statistical index from the relationship structure D2 and the abnormality-time relationship structure D36, respectively. The relationship change estimation unit 340 selects statistical indices that are meaningful in comparing statistics as statistical indices to be extracted. For example, the relationship change estimation unit 340 determines that the statistical indices calculated based on the p-value of the hypothesis test are not statistical indices that are meaningful in comparing statistics, and selects the other statistical indices as statistical indices that are meaningful in comparing statistics. The relationship change estimation unit 340 can estimate the breakdown of the relationship, in other words, the change in the relationship, by comparing the statistical amount corresponding to the correlation, that is, the magnitude relationship of the correlation coefficient.

[0154] Next, the relationship change estimation unit 340 calculates the amount of change d(k) based on the elements a(k) of the relationship change estimation association structure and the elements a'(k) of the abnormality-time relationship structure for relationship change estimation. Then, the relationship change estimation unit 340 generates an association structure change amount, which is information in which the amount of change d(k) is represented in a matrix and has the calculated amount of change d(k) as an element. Here, the elements a(k)ij, a'(k)ij and the amount of change d(k)ij are all real values. The relationship change estimation unit 340 may calculate the amount of change d(k)ij, which is an element of the amount of change in the relationship structure, by using, for example, the absolute value of the difference between the absolute value of the element a(k)ij of the relationship change estimation association structure and the absolute value of the element a'(k)ij of the relationship change estimation abnormality-time association structure, or may calculate the amount of change d(k)ij by using the absolute value of the difference between the element a(k)ij of the relationship change estimation association structure and the element a'(k)ij of the relationship change estimation abnormality-time association structure.

[0155] Next, the relationship change estimation unit 340 acquires the anomaly detection sensor information D3 from the data storage unit 20, and calculates the relationship change degree for each sensor 300, more specifically, for each anomaly detection sensor, based on the change amount d(k)ij, which is an element of the calculated associated structural change amount, and the anomaly detection sensor information D3. Here, both the change amount d(k)ij and the relationship change degree are real values. The relationship change estimation unit 340 calculates the relationship change degree corresponding to the anomaly detection sensor Xi, which is the i-th sensor 300 among the n sensors 300, for example, by using the average of elements other than the i-th column in the i-th row of the associated structure change amount and corresponding to the sensor 300 (i.e., the anomaly detection sensor) included in the anomaly detection sensor information D3. Note that this is merely an example, and the relationship change estimation unit 340 may calculate the relationship change degree corresponding to the anomaly detection sensor Xi, which is the i-th sensor 300 among the n sensors 300, by using the average of elements other than the i-th column in the i-th row of the associated structure change amount.

[0156] Next, the relationship change estimation unit 340 assigns a relationship change order corresponding to the anomaly detection sensor based on the calculated relationship change degree corresponding to the anomaly detection sensor. Here, both the relationship change degree and the relationship change order are real values. The relationship change estimation unit 340 assigns the relationship change order so that the corresponding relationship change orders are in ascending order, for example, starting from the anomaly detection sensor with the largest relationship change degree value. However, when the relationship change degrees corresponding to multiple anomaly detection sensors are equal, the relationship change estimation unit 340 assigns the same relationship change order to the multiple anomaly detection sensors. After allocating the relationship change order, the relationship change estimation unit 340 generates a relationship change order estimation result D37. The relationship change order estimation result is information in which information indicating the anomaly detection sensor, a relationship change degree, and a relationship change order are associated with each other. The relationship change estimation unit 340 stores the relationship change order estimation result D37 in the data storage unit 20.

[0157] The relationship change order estimation process performed by the relationship change estimation unit 340 as described above will be described below with reference to the drawings, giving a specific example. FIG. 22 is a diagram for explaining the concept of an example of a relationship change order estimation process performed by the relationship change estimation unit 340 based on the learned relationship structure D2 and the relationship structure D36 in the abnormal state when the abnormality factor estimating device 100 according to the first embodiment is configured to include the relationship change estimation unit 340. Here, as an example, the number of sensors 300 is 4, and the sensors 300 are represented as sensors Xn (n=1, 2, 3, 4). Also, as an example, of the sensors X1, X2, X3, and X4, the sensors X1, X2, and X3 are anomaly detection sensors.

[0158] First, the relationship change estimation unit 340 acquires the relationship structure D2 and the abnormal-time related structure D36 from the data storage unit 20, and acquires the relationship change estimation related structure D2A and the abnormal-time related structure D36A from the relationship structure D2 and the abnormal-time related structure D36. Note that the relationship structure D2 and the abnormal-time related structure D36 are omitted from FIG. 22. In FIG. 22, the relationship change estimation association structure D2A is shown as a two-dimensional matrix in which only a portion corresponding to the k-th statistical indicator is extracted from the association structure D2. The relationship change estimation association structure D2A is expressed as a two-dimensional matrix in which the first dimension is the quantity 4 of the sensors Xn and the second dimension is the quantity 4 of the sensors Xn. The statistical indicator is a non-directional statistical indicator, and the type of the statistical indicator is correlation. Note that, as shown in FIG. 22, the relationship change estimation anomaly-time association structure D36A and the association structure change amount D39 have the same data structure as the relationship change estimation association structure D2A.

[0159] 22, as an example, the amount of change d(k)ij, which is an element of the association structure change amount D39, is the absolute value of the difference between the absolute value of the element a(k)ij of the association structure for estimating a relationship change D2A and the absolute value of the element a'(k)ij of the association structure for estimating a relationship change at an abnormal time D36A. For example, the relationship change estimation unit 340 calculates the amount of change between the sensor X1 and the sensor X2 as ||a(k)12|-|a'(k)12||=d(k)12 based on the element a(k)12 in the first row and second column of the association structure for estimating a relationship change D2A and the element a'(k)12 in the first row and second column of the association structure for estimating a relationship change at an abnormal time D36A. Here, |·| represents the absolute value. 22, for example, the relationship change estimation unit 340 calculates the relationship change degree dn corresponding to the anomaly detection sensor Xn by using the average of elements in the nth row of the associated structure change amount D39 other than the nth column and corresponding to the sensor 300 (i.e., the anomaly detection sensor) included in the anomaly detection sensor information D3. Therefore, the relationship change estimation unit 340 calculates the relationship change degree d1 corresponding to the anomaly detection sensor X1, for example, as the average of elements d(k)nn (here, n=1, 2, 3) in the 1st row of the associated structure change amount D39 other than the 1st column and corresponding to the anomaly detection sensors X1, X2, and X3. Specifically, the relationship change estimation unit 340 calculates the relationship change degree d1 of the anomaly detection sensor X1 as the average of the change amount d(k)12, which is the element in the 1st row and 2nd column of the associated structure change amount D39 corresponding to the anomaly detection sensors X2 and X3, and the change amount d(k)13, which is the element in the 1st row and 3rd column of the associated structure change amount D39. The relationship change estimation unit 340 similarly calculates relationship change degrees d2 and d3 corresponding to the sensors X2 and X3, respectively. Based on the calculated relationship change degrees d1, d2, and d3, the relationship change estimation unit 340 assigns corresponding relationship change orders o1, o2, and o3 to the sensors X1, X2, and X3, respectively. Then, the relationship change estimation unit 340 stores the relationship change order estimation result D37 in the data storage unit 20.

[0160] The abnormality factor estimation unit 60 acquires the abnormality detection order estimation result D5, the abnormality propagation order estimation result D6, and the relationship change order estimation result D37 from the data storage unit 20, and performs an abnormality factor estimation process taking into consideration the relationship change order based on the abnormality detection order estimation result D5, the abnormality propagation order estimation result D6, and the relationship change order estimation result D37. In detail, performing an abnormality factor estimation process taking into consideration the relationship change order based on the abnormality detection order estimation result D5, the abnormality propagation order estimation result D6, and the relationship change order estimation result D37 means that the abnormality factor estimation unit 60 calculates corresponding abnormality factor scores from the abnormality detection order included in the abnormality detection order estimation result D5, the abnormality propagation order included in the abnormality propagation order estimation result D6, and the relationship change order included in the relationship change order estimation result D37, and estimates the abnormality factor order based on the calculated abnormality factor scores. The abnormality factor estimation unit 60 may calculate the abnormality factor score from the abnormality detection order, the abnormality propagation order, and the relationship change order in a manner similar to that for calculating the abnormality factor score from the abnormality detection order and the abnormality propagation order. The abnormality factor estimating unit 60 stores the abnormality factor order estimation result D40, which takes into account the relationship change order, in the data storage unit 20.

[0161] The abnormality factor estimation result output unit 70 acquires the abnormality factor order estimation result D40, the abnormality detection order estimation result D5, the abnormality propagation order estimation result D6, and the relationship change order estimation result D37 from the data storage unit 20, and outputs information related to the estimation result of the abnormality factors by the abnormality factor estimation unit 60. In detail, the abnormality factor estimation result output unit 70 outputs, to the display device 400, abnormality factor estimation result display information for displaying an abnormality factor estimation result screen showing information related to the estimation result of the abnormality factors estimated by the abnormality factor estimation unit 60, based on the abnormality factor order estimation result D40, the abnormality detection order estimation result D5, the abnormality propagation order estimation result D6, and the relationship change order estimation result D37.

[0162] Note that, when the abnormality factor estimating device 100 is configured to include a relationship change estimation unit 340 as shown in FIG. 21 , in the operation of the abnormality factor estimating device 100 described using the flowchart of FIG. 13 , the relationship change estimation unit 340 estimates a change in the relationship between the sensor data before the processing of step ST5 is performed, and stores the relationship change order estimation result D37 in the data storage unit 20. In step ST5, the abnormality factor estimation unit 60 acquires the abnormality detection order estimation result D5, the abnormality propagation order estimation result D6, and the relationship change order estimation result D37 from the data storage unit 20, and performs abnormality factor estimation processing taking into account the relationship change order based on the abnormality detection order estimation result D5, the abnormality propagation order estimation result D6, and the relationship change order estimation result D37.

[0163] In this way, the abnormality factor estimation device 100 includes a relationship change estimation unit 340 that compares the relationship structure D2 (estimated structure) with the relationship structure D36 at the time of abnormality occurrence based on the relationship structure D2 (estimated structure), the relationship structure D36 at the time of abnormality occurrence, and the abnormality detection sensor information D3 to estimate a change in the relationship between the sensor data, and the abnormality factor estimation unit 60 is configured to estimate the abnormality factor by taking into consideration the change in the relationship between the sensor data estimated by the relationship change estimation unit 340 based on the abnormality detection order estimated by the abnormality detection order estimation unit 40 and the abnormality propagation order estimated by the abnormality propagation path tracking unit 50, thereby improving the reliability of the abnormality factor order estimation result D7 and enabling the estimation of the abnormality factor to be performed precisely. By including the relationship change estimation unit 340 in the abnormality factor estimation device 100, a criterion for estimating the abnormality factor in the abnormality factor estimation device 100 is added.

[0164] <Modification> In the above first embodiment, the abnormality factor estimation device 100 "estimates the cause of an abnormality" means estimating an abnormality factor score indicating the degree of likelihood of an abnormality source and an abnormality factor order based on the abnormality factor score for each sensor 300, and generating information on the abnormality factor score and the abnormality factor order. Furthermore, the abnormality factor estimation device 100 "estimates the cause of an abnormality" may also include estimating an abnormality factor score for each device and an abnormality factor order based on the abnormality factor score. FIG. 23 is a diagram showing an example of the configuration of the abnormality factor estimating device 100 according to the first embodiment, which includes an abnormality factor device inference unit 350 and is configured to estimate an abnormality factor on an equipment basis. 23, the abnormality factor estimation device 100 includes a sensor data acquisition unit 10, an abnormality detection unit 30, an abnormality detection order estimation unit 40, an abnormality propagation path tracking unit 50, an abnormality factor estimation unit 60, and a control unit in addition to the abnormality factor device estimation unit 350, the data storage unit 20, and the abnormality factor estimation result output unit 70. However, the abnormality factor estimation device 100 does not necessarily have to include the abnormality factor estimation result output unit 70. In addition, although not shown in FIG. 23 for simplicity of explanation, the abnormality factor estimation device 100 is connected to a learning device 200.

[0165] The abnormality factor device inference unit 350 acquires device-specific sensor information D41. The abnormality factor device inference unit 350 also acquires the abnormality detection order estimation result D5 and the abnormality propagation order estimation result D6 from the data storage unit 20. The abnormality factor device inference unit 350 performs an apparatus-specific abnormality factor inference process that infers the cause of an abnormality on an apparatus-by-apparatus basis, based on the device-specific sensor information D41, the abnormality detection order estimation result D5, and the abnormality propagation order estimation result D6. The device-specific sensor information D41 is table data indicating which device the sensor 300 is provided in. For example, an operator or the like operates an input device such as a mouse or a keyboard to input the device-specific sensor information D41, and the abnormality factor device estimation unit 350 receives the input device-specific sensor information D41 to acquire the device-specific sensor information D41. For example, in the device-attached sensor information D41, information indicating a device is associated with information indicating the sensor 300 provided in the device.

[0166] The equipment-specific abnormality factor inference process performed by the abnormality factor equipment inference unit 350 will be described in detail below. First, the abnormality factor device estimation unit 350 acquires the device-specific sensor information D41, the abnormality detection order estimation result D5, and the abnormality propagation order estimation result D6. The abnormality factor device estimation unit 350 converts the abnormality detection order estimation result D5 into an abnormality detection order estimation result on an equipment basis (hereinafter referred to as an "equipment abnormality detection order estimation result") D42 based on the device attached sensor information D41 and the abnormality detection order estimation result D5. Specifically, for a certain device U among U devices (U is an integer), the abnormality factor device estimation unit 350 compares information indicating the sensor 300 associated in the device-attached sensor information D41 with information indicating the sensor 300 included in the abnormality detection order estimation result D5. Then, the abnormality factor device estimation unit 350 acquires the abnormality detection order associated with the information indicating the associated sensor 300 in the abnormality detection order estimation result D5, and calculates the abnormality detection order total value. Here, the abnormality factor device estimation unit 350 uses, for example, a weighted average of the abnormality detection orders associated with the information indicating the associated sensor 300 to set a representative value as the abnormality detection order total value. In addition, the abnormality factor device estimation unit 350 may set, for example, a representative value such as a minimum value or a maximum value of the abnormality detection orders associated with the information indicating the associated sensor 300 as the abnormality detection order total value. Then, the abnormality factor device inference unit 350 assigns an abnormality detection order for each device (hereinafter referred to as "device abnormality detection order") to the device U based on the calculated abnormality detection order tabulation value. Here, the device abnormality detection order ouU is a real value. The abnormality factor device inference unit 350 assigns the device abnormality detection order to the device u, for example, in ascending order of the abnormality detection order tabulation value, so that the corresponding device abnormality detection orders are in ascending order. However, when the calculated abnormality detection order tabulation values ​​are equal for multiple devices, the abnormality factor device inference unit 350 assigns the same device abnormality detection order to the multiple devices. The abnormality factor equipment estimation unit 350 generates an equipment abnormality detection order estimation result D42, which is information in which information indicating the equipment, the abnormality detection order aggregate value, and the equipment abnormality detection order are associated with each other on an equipment basis, and stores the equipment abnormality detection order estimation result D42 in the data storage unit 20.

[0167] In addition, the abnormality factor device estimation unit 350 converts the abnormality propagation order estimation result D6 into an abnormality propagation order estimation result for each device (hereinafter referred to as "device abnormality propagation order estimation result") D43 based on the device-specific sensor information D41 and the abnormality propagation order estimation result D6. Specifically, assuming that U devices are represented by devices U (U=1, 2, . . . , U), the abnormality factor device estimation unit 350 compares, for a certain device U, information indicating the sensor 300 associated in the device-attached sensor information D41 with information indicating the sensor 300 included in the abnormality propagation order estimation result D6. Then, the abnormality factor device estimation unit 350 acquires the abnormality propagation order associated with the information indicating the associated sensor 300 in the abnormality propagation order estimation result D6, and calculates the abnormality propagation order tally value. Here, the abnormality factor device estimation unit 350 uses, for example, a weighted average of the abnormality propagation order associated with the information indicating the associated sensor 300, and sets the representative value as the abnormality propagation order tally value. In addition, the abnormality factor device estimation unit 350 may set, for example, a representative value such as a minimum value or a maximum value of the abnormality propagation order associated with the information indicating the associated sensor 300 as the abnormality propagation order tally value. Then, the abnormality factor device estimation unit 350 assigns an abnormality propagation order for each device U (hereinafter referred to as "device abnormality propagation order") based on the calculated abnormality propagation order tabulation value. Here, the abnormality detection order is a real value. The abnormality factor device estimation unit 350 assigns the device abnormality propagation order to the device U in ascending order of the corresponding device abnormality propagation order, for example, from the smallest abnormality propagation order tabulation value. However, when the calculated abnormality propagation order tabulation values ​​are equal for multiple devices, the abnormality factor device estimation unit 350 assigns the same device abnormality propagation order to the multiple devices. The abnormality factor device estimation unit 350 generates an equipment abnormality propagation order estimation result D43, which is information in which information indicating the device, an abnormality detection device flag, an abnormality propagation order aggregate value, and an equipment abnormality propagation order are associated with each other for each device, and stores the equipment abnormality propagation order estimation result D43 in the data storage unit 20. The abnormality detection device flag indicates, for each device, whether or not there is an abnormality detection sensor among the sensors 300 provided in the device. The abnormality detection device flag is a Boolean value.

[0168] As described above, in the first embodiment, the appliance may be provided with a plurality of sensors 300. For example, if there is at least one anomaly detection sensor among the plurality of sensors 300 provided in a certain appliance U, the anomaly factor appliance estimation unit 350 sets the anomaly detection appliance flag corresponding to the certain appliance U to (True) in the appliance anomaly propagation order estimation result D43. That is, for example, if a certain appliance U is provided with a plurality of sensors 300 and there is one or more anomaly detection sensors among the plurality of sensors 300, the anomaly factor appliance estimation unit 350 sets the anomaly detection appliance flag corresponding to the certain appliance U to (True). On the other hand, for example, if a certain appliance U is provided with a plurality of sensors 300 and there is no anomaly detection sensor among the plurality of sensors 300, the anomaly factor appliance estimation unit 350 sets the anomaly detection appliance flag corresponding to the certain appliance U to (False).

[0169] Furthermore, the abnormality factor device inference unit 350 infers the cause of the abnormality for each device based on the generated device abnormality detection order estimation result D42 and device abnormality propagation order estimation result D43. Specifically, the abnormality factor device estimation unit 350 calculates an appliance abnormality factor score for each appliance from the appliance abnormality detection order set in the appliance abnormality detection order estimation result D42 and the appliance abnormality propagation order set in the appliance abnormality propagation order estimation result D43. Here, the appliance abnormality factor score is a real value. For example, the abnormality factor device estimation unit 350 uses a weighted average of the device abnormality detection order and the device abnormality propagation order for each device and sets a representative value as the device abnormality factor score. Also, for example, the abnormality factor device estimation unit 350 may set a representative value such as a maximum value or a minimum value of the device abnormality detection order and the device abnormality propagation order for each device as the device abnormality factor score. Note that when only one of the device abnormality detection order and the device abnormality propagation order is set, the abnormality factor device estimation unit 350 may use the set value as the device abnormality factor score as it is. In this case, the abnormality factor device estimation unit 350 may weight the device abnormality factor score taking into consideration that only one order is set.

[0170] Then, the abnormality factor device estimation unit 350 assigns an abnormality factor order for each device (hereinafter referred to as "device abnormality factor order") based on the calculated device abnormality factor score for each device. Here, the device abnormality factor order is a real value. For example, the abnormality factor device estimation unit 350 assigns the device abnormality factor order to the device U in ascending order of the corresponding device abnormality factor order from the device with the smallest corresponding device abnormality factor score value. However, if the calculated device abnormality factor scores are equal for multiple devices U, the abnormality factor device estimation unit 350 assigns the same device abnormality factor order to the multiple devices U. The abnormality factor device estimation unit 350 generates an appliance abnormality factor order estimation result D44, which is information in which information indicating the appliance, an abnormality detection appliance flag, an appliance abnormality factor score, and an appliance abnormality factor order are associated with each other for each appliance, and stores the appliance abnormality factor order estimation result D44 in the data storage unit 20. Note that the abnormality factor device estimation unit 350 may set, for the appliance U, the value of the abnormality detection appliance flag set in association with the information indicating the appliance U in the appliance abnormality propagation order estimation result D43, as the abnormality detection appliance flag associated with the appliance abnormality factor order estimation result D44.

[0171] The above-mentioned equipment-specific abnormality factor inference process by the abnormality factor equipment inference unit 350 will be described with reference to the drawings and a specific example. FIG. 24 is a diagram illustrating a concept of an example of an apparatus-by-apparatus anomaly factor estimation process that is performed by the anomaly factor equipment estimation unit 350 when the anomaly factor estimation device 100 according to the first embodiment is configured to include the anomaly factor equipment estimation unit 350, based on the equipment-attached sensor information D41, the anomaly detection order estimation result D5, and the anomaly propagation order estimation result D6, to estimate the cause of an anomaly on an apparatus-by-apparatus basis. Here, as an example, the quantity of sensors 300 is 6, and the sensors 300 are represented by sensors Xn (n=1 to 6). As another example, the quantity of devices is 3, and the devices are represented by devices U (U=1, 2, 3). Also, it is assumed that device 1 is provided with sensors X1 and X2, device 2 is provided with sensors X3, X4, and X5, and device 3 is provided with sensor X6.

[0172] The abnormality factor device inference unit 350 calculates the abnormality detection order tabulation value suU (shown by D42B in FIG. 24) for each device U, for example, by using an average, based on the abnormality detection order on (shown by D5C in FIG. 24) set in the abnormality detection order estimation result D5 and the device-attached sensor information D41. For example, in device 1, the sensors Xn added to device 1 are sensor X1 and sensor X2, so the abnormality factor device inference unit 350 sets the abnormality detection order tabulation value su1 corresponding to device 1 to the average value of the abnormality detection order o1 and the abnormality detection order o2. Also, for example, in device 2, the sensors Xn added to device 2 are sensor X3, sensor X4, and sensor X5. However, of sensors X3, X4, and X5, the only abnormality detection sensor, in other words, the sensor Xn included in the abnormality detection order estimation result D5, is sensor X4. Therefore, the abnormality factor device inference unit 350 sets the abnormality detection order o4 corresponding to the sensor X4 as the abnormality detection order tabulation value su2 corresponding to the device 2. Similarly, the abnormality factor device inference unit 350 sets the abnormality detection order o6 corresponding to the sensor X6 as the abnormality detection order tabulation value su6 corresponding to the device 3.

[0173] Next, the abnormality factor device inference unit 350 assigns an appliance abnormality detection order ouU to each appliance based on the abnormality detection order total value suU calculated for each appliance. Specifically, the abnormality factor device inference unit 350 assigns appliance abnormality detection orders ou1, ou2, ou3 to appliances 1, 2, 3, respectively, based on the calculated abnormality detection order total values ​​su1, su2, su3. Then, the abnormality factor equipment estimation unit 350 generates an equipment abnormality detection order estimation result D42 in which information indicating equipment 1, 2, and 3 (shown by D42A in FIG. 24), the abnormality detection order aggregate values ​​su1, su2, and su3 (shown by D42B in FIG. 24), and the equipment abnormality detection orders ou1, ou2, and ou3 (shown by D42C in FIG. 24) are associated with each other, and stores the equipment abnormality detection order estimation result D42 in the data storage unit 20.

[0174] Furthermore, the abnormality factor device estimation unit 350 calculates the abnormality propagation order tabulation value suU (shown by D43C in FIG. 24) for each device based on the abnormality propagation order on (shown by D6C in FIG. 24) set in the abnormality propagation order estimation result D6 and the device-attached sensor information D41, for example, by using an average. For example, in device 1, the sensor Xn added to device 1 is sensor X1 and sensor X2, so the abnormality factor device estimation unit 350 sets the abnormality propagation order tabulation value su1 corresponding to device 1 to the average value of the abnormality propagation order o1 and the abnormality propagation order o2. For example, in device 2, the sensor Xi added to device 2 is sensor X3, sensor X4, and sensor X5, so the abnormality factor device estimation unit 350 sets the abnormality propagation order tabulation value su2 corresponding to device 2 to the average value of the abnormality propagation order o3, the abnormality propagation order o4, and the abnormality propagation order o5. Similarly, the abnormality factor device inference unit 350 sets the abnormality propagation order o6 corresponding to the sensor X6 as the abnormality propagation order aggregate value su6 corresponding to the device 3 without any change.

[0175] Next, the abnormality factor device inference unit 350 assigns an equipment abnormality propagation order ouU based on the abnormality propagation order total value suU calculated for each device. Specifically, the abnormality factor device inference unit 350 assigns equipment abnormality propagation orders ou1, ou2, and ou3 corresponding to the devices 1, 2, and 3, respectively, based on the calculated abnormality propagation order total values ​​su1, su2, and su3. Then, the abnormality cause equipment estimation unit 350 generates an equipment abnormality propagation order estimation result D43 in which information indicating equipment 1, 2, and 3 (shown by D43A in FIG. 24), the abnormality detection equipment flag (shown by D43B in FIG. 24), the abnormality propagation order summary values ​​su1, su2, and su3 (shown by D43C in FIG. 24), and the equipment abnormality propagation orders ou1, ou2, and ou3 (shown by D43D in FIG. 24) are associated with each other, and stores the equipment abnormality propagation order estimation result D43 in the data storage unit 20.

[0176] Furthermore, the abnormality factor device estimation unit 350 calculates an appliance abnormality factor score suU for each appliance by using an average based on the appliance abnormality detection order ouU set in the appliance abnormality detection order estimation result D42 and the appliance abnormality propagation order ouU set in the appliance abnormality propagation order estimation result D43. For example, for appliance 1, the abnormality factor device estimation unit 350 calculates the average value of the appliance abnormality detection order ou1 and the appliance abnormality propagation order ou1 of appliance 1 as the appliance abnormality factor score su1. Then, the abnormality factor device inference unit 350 assigns an appliance abnormality factor order ouU to each appliance based on the appliance abnormality factor score suU. For example, the abnormality factor device inference unit 350 assigns corresponding appliance abnormality factor orders ou1, ou2, and ou3 to appliances 1, 2, and 3, respectively, based on the calculated appliance abnormality factor scores su1, su2, and su3. The abnormality factor equipment estimation unit 350 generates an equipment abnormality factor order estimation result D44, which is information in which information indicating equipment 1, 2, and 3 (shown by D44A in FIG. 24), an abnormality detection equipment flag (shown by D44B in FIG. 24), equipment abnormality factor scores su1, su2, su3 (shown by D44C in FIG. 24), and equipment abnormality factor orders ou1, ou2, ou3 (shown by D44D in FIG. 24) are associated with each other, and stores the equipment abnormality factor order estimation result D44 in the data storage unit 20.

[0177] As described above, when the abnormality factor equipment estimation unit 350 stores the equipment abnormality detection order estimation result D42, the equipment abnormality propagation order estimation result D43, and the equipment abnormality factor order estimation result D44 in the data storage unit 20, the abnormality factor estimation result output unit 70 acquires the equipment abnormality detection order estimation result D42, the equipment abnormality propagation order estimation result D43, and the equipment abnormality factor order estimation result D44 output by the abnormality factor equipment estimation unit 350 via the data storage unit 20, and outputs information related to the estimation result of the cause of abnormality for each equipment based on the equipment abnormality detection order estimation result D42, the equipment abnormality propagation order estimation result D43, and the equipment abnormality factor order estimation result D44. In detail, the abnormality factor estimation result output section 70 outputs information (hereinafter referred to as "abnormality factor equipment estimation result display information") for displaying a screen showing information on the estimation result of the cause of abnormality for each equipment by the abnormality factor equipment estimation section 350 (hereinafter referred to as "abnormality factor equipment estimation result display information") to the display device 400 based on the equipment abnormality detection order estimation result D42, the equipment abnormality propagation order estimation result D43, and the equipment abnormality factor order estimation result D44.

[0178] FIG. 25 is a diagram showing an example of an abnormality factor device inference result screen displayed on the display device 400 by the abnormality factor inference result output unit 70 in the first embodiment. FIG. 25 shows an example of an abnormality factor device estimation result screen in the case where the quantity of devices is set to 3 (device U, U=1 to 3). In FIG. 25, the abnormality factor device estimation result screen is shown as "D48-1." For example, as shown in FIG. 25, the abnormality factor equipment inference result screen has ten display frames: display frame D48A, display frame D48B, display frame D48C, display frame D48D, display frame D48E, display frame D48F, display frame D48G, display frame D48H, display frame D48I, and display frame D48J. The abnormality factor estimation result output unit 70 displays, for example, an abnormality factor equipment estimation result list in which information on the estimation results of the abnormality factors is listed for each equipment on an abnormality factor equipment estimation result screen. The abnormality factor equipment estimation result list is, for example, a list in which, for each equipment U, information indicating the equipment U, information indicating the abnormality detection equipment flag, equipment abnormality detection order, equipment abnormality propagation order, equipment abnormality factor score, and equipment abnormality factor order are displayed in association with each other. On the abnormality factor equipment estimation result screen shown in FIG. 25, the abnormality factor equipment estimation result list is indicated by "D48-1a".

[0179] The abnormality factor estimation result output unit 70 displays information indicating the equipment U of the equipment abnormality factor order estimation result D44 in the display frame D48A, displays information indicating the abnormality detection equipment flag of the equipment abnormality propagation order estimation result D43 in the display frame D48B, displays the equipment abnormality detection order of the equipment abnormality detection order estimation result D42 in the display frame D48C, displays the equipment abnormality propagation order of the equipment abnormality propagation order estimation result D43 in the display frame D48D, displays the equipment abnormality factor score of the equipment abnormality factor order estimation result D44 in the display frame D48E, and displays the equipment abnormality factors of the equipment abnormality factor order estimation result D44 in the display frame D48F. The order is displayed in display frame D48F, sort buttons for rearranging the order of the abnormality factor device estimation result list in ascending order based on the device abnormality detection order of the device abnormality detection order estimation result D42, the device abnormality propagation order of the device abnormality propagation order estimation result D43, and the device abnormality factor order of the device abnormality factor order estimation result D44 are displayed in display frames D48H, D48I, and D48J, respectively, and a check box for receiving an instruction to display only the abnormality detected devices is displayed in display frame D48G, and abnormality factor device estimation result display information is output to the display device 400. As a result, the display device 400 displays an abnormality factor device estimation result screen as shown in FIG.

[0180] The anomaly factor device estimation result screen as shown in FIG. 25 is the anomaly factor estimation result screen already explained using FIG. 8, displayed on an equipment basis, and the functions of the sort buttons and check boxes are similar to those of the sort buttons and check boxes already explained using FIG. 8, so duplicated explanations will be omitted.

[0181] In this case, in the operation of the abnormality factor estimation device 100 described using the flowchart in FIG. 13, the abnormality factor equipment estimation unit 350 performs equipment-by-equipment abnormality factor estimation processing to estimate the cause of the abnormality on an equipment-by-equipment basis based on the equipment-attached sensor information D41, the abnormality detection order estimation result D5, and the abnormality propagation order estimation result D6 before or after the processing of step ST5. Also, for example, the abnormality cause inference result output unit 70 may be able to select whether to output information regarding the inference result of the abnormality cause on a sensor-by-sensor basis, as described in the above first embodiment, or to output information regarding the inference result of the abnormality cause on an equipment-by-equipment basis.

[0182] In this way, the abnormality factor estimation device 100 can be configured to include an abnormality factor device estimation unit 350 that estimates the cause of an abnormality for each device based on the device-attached sensor information, the abnormality detection order estimated by the abnormality detection order estimation unit 40, and the abnormality propagation order estimated by the abnormality propagation path tracing unit 50. In this way, the abnormality factor estimation device 100 can enable the operator to efficiently identify the device that caused the abnormality. In addition, the abnormality factor estimation device 100 can enable the operator to efficiently understand the order in which the devices in which an abnormality has occurred should be inspected.

[0183] <Modification> In the above-described first embodiment, the abnormality factor estimating device 100 may be configured to include an associated structure graph output unit 360 that outputs information for displaying a graph related to the associated structure D2 stored in the data storage unit 20 (hereinafter referred to as “associated structure graph display information”) to the display device 400. FIG. 26 is a diagram showing an example of the configuration of the abnormality factor estimating device 100 according to the first embodiment, which includes an associated structure graph output unit 360 and is configured to output associated structure graph display information to a display device 400. 26, the abnormality factor estimation device 100 includes a sensor data acquisition unit 10, an abnormality detection unit 30, an abnormality detection order estimation unit 40, an abnormality propagation path tracking unit 50, an abnormality factor estimation unit 60, an abnormality factor estimation result output unit 70, and a control unit, in addition to the association structure graph output unit 360 and the data storage unit 20. However, it is not essential that the abnormality factor estimation device 100 includes the abnormality factor estimation result output unit 70. In addition, although not shown in FIG. 26 for simplicity of explanation, the abnormality factor estimation device 100 is connected to a learning device 200.

[0184] The associated structure graph output unit 360 outputs associated structure graph display information for displaying a graph associated with the associated structure D2 to the display device 400 based on the associated structure D2, the anomaly detection sensor information D3, and the anomaly cause order estimation result D7 stored in the data storage unit 20. The graph associated with the associated structure D2 is, for example, a graph in which the associated structure D2, the anomaly detection sensor, and the estimation result of the anomaly cause are associated with each other. Based on the associated structure graph display information output from the associated structure graph output unit 360, the display device 400 displays a screen showing a graph related to the associated structure D2 (hereinafter referred to as a "graph screen"). The associated structure graph output unit 360 may be provided in a device, such as the display device 400, that is external to the abnormality factor estimating apparatus 100 and connected to the abnormality factor estimating apparatus 100 via a wired or wireless signal line.

[0185] When the abnormality factor estimation device 100 is configured to include an associated structure graph output unit 360, an example of a graph screen that is displayed on the display device 400 by outputting associated structure graph display information from the associated structure graph output unit 360 will be described with reference to the drawings. FIG. 27 is a diagram for explaining an example of a graph screen displayed on the display device 400 by outputting associated structure graph display information from the associated structure graph output unit 360 when the abnormality factor estimating device 100 is configured to include the associated structure graph output unit 360 in the first embodiment.

[0186] Here, as an example, the number of sensors 300 is 6, and the sensors 300 are represented as sensors Xn (n=1 to 6). 27 is an example of a graph screen displayed based on the associated structure graph display information output by the associated structure graph output unit 360 when the contents of the associated structure D2 stored in the data storage unit 20 are as shown in FIG. 28, the contents of the abnormality detection sensor information D3 are as shown in FIG. 29, and the abnormality factor order estimation result D7 is as shown in FIG. 30.

[0187] When outputting information for displaying an association structure graph, the association structure graph output unit 360 determines, for each statistical indicator, whether each element of the association structure D2 has a dependent relationship or does not have a dependent relationship, and converts the association structure D2 into an association structure in which the elements include information indicating the presence or absence of a dependent relationship (hereinafter referred to as the "association structure after dependency determination"). For example, for each statistical indicator, the association structure graph output unit 360 determines that each element of the association structure D2 is an element with a dependency relationship if the statistical indicator is equal to or greater than a preset threshold for dependency selection, and determines that the element is an element without a dependency relationship if the statistical indicator is less than the threshold for dependency selection.The association structure graph output unit 360 then generates a post-dependency relationship determination association structure that is indicated by a matrix in which elements with a dependency relationship are assigned a value of "1" and elements without a dependency relationship are assigned a value of "0". In FIG. 28, the dependency relationship determination degree association structure for each statistical indicator (shown as D2R in FIG. 28) is illustrated together with the association structure D2.

[0188] In Fig. 27, the graph screen is indicated by D45. On the graph screen, for example, an association structure graph (indicated by D45-1 in Fig. 27) which is a directed graph in which the association structure D2 is expressed with the sensors Xn as nodes and the dependency relationships between the sensors Xn as edges is displayed.

[0189] In addition to the association structure graph, the graph screen displays a screen (hereinafter referred to as an "index specification screen") (shown by D45I in FIG. 27) that displays check boxes (hereinafter referred to as "index specification check boxes") for accepting the specification of the type of statistical index for which the corresponding element in the association structure graph is to be displayed. Here, since there are three types of statistical indexes, the index specification screen shown in FIG. 27 displays an index specification check box corresponding to "statistical index 1" for accepting the specification of the first type of statistical index, an index specification check box corresponding to "statistical index 2" for accepting the specification of the second type of statistical index, and an index specification check box corresponding to "statistical index 3" for accepting the specification of the third type of statistical index. For example, the operator specifies the statistical index for which the correlation (edge) between the corresponding sensors Xn is to be displayed by checking the index specification check box. FIG. 27 shows a state in which the index specification check box corresponding to "statistical index 1" and the index specification check box corresponding to "statistical index 2" are checked on the index specification screen, in other words, a state in which the first type of statistical index and the second type of statistical index are specified. Therefore, only edges corresponding to the first and second types of statistical indicators, which are statistical indicators with their indicator designation checkboxes checked, are displayed on the association structure graph. For example, as shown in Fig. 27, edges corresponding to each statistical indicator are displayed with different line types so that it is clear which statistical indicator each edge corresponds to. In the association structure graph shown in Fig. 27, edges corresponding to a first type of statistical indicator are displayed with solid arrows (see, for example, D45G), and edges corresponding to a second type of statistical indicator are displayed with dotted arrows (see, for example, D45H). Note that this is merely an example, and for example, edges corresponding to each statistical indicator may be displayed with different arrow colors.

[0190] Also, on the graph screen, a screen (hereinafter referred to as a "node condition specification screen") (shown by D45J in FIG. 27) displaying check boxes (hereinafter referred to as "node condition specification check boxes") for specifying display conditions related to nodes (hereinafter referred to as "node display conditions") is displayed. The node display conditions are set in advance. In the node condition specification screen shown in FIG. 27, three conditions, "display only anomaly detection sensors", "highlight anomaly detection sensors", and "display order of anomaly factors", are set as node display conditions. For example, an operator specifies node display conditions by checking the node condition specification check boxes. FIG. 27 shows a state in which the node condition specification check boxes corresponding to "highlight anomaly detection sensors" and "display order of anomaly factors" are checked on the node condition specification screen. Therefore, in the association structure graph, the nodes corresponding to the anomaly detection sensors X1, X4, X5, and X6 (shown as D45A, D45D, D45E, and D45F in Fig. 27) are displayed as filled in. Note that, although the anomaly detection sensors are highlighted here by displaying the nodes corresponding to the anomaly detection sensors as filled in, the method of highlighting the anomaly detection sensors is not limited to this, and the nodes corresponding to the anomaly detection sensors may be highlighted in other ways.

[0191] In addition, in the association structure graph, the order of abnormality factors is displayed on the node corresponding to the sensor Xn. In the association structure graph shown in Fig. 27, the order of abnormality factors is displayed as "rank1", "rank2", "rank3", "rank4", or "rank5". Note that, although the order of abnormality factors is displayed as "rank△" here, the method of displaying the order of abnormality factors is not limited to this, and it is sufficient if the order of abnormality factors is displayed in a way that makes it possible to understand it. In this example, the three node display conditions are set to "display only anomaly detection sensors," "highlight anomaly detection sensors," and "display order of anomaly causes." However, this is merely an example, and other conditions may be set to the node display conditions.

[0192] In the association structure graph, for example, a sensor name is displayed at each node so that an operator checking the graph screen can identify the sensor Xn. In Fig. 27, the sensor names "X1", "X2", "X3", "X4", "X5", and "X6" are displayed. For example, an edge is displayed when a statistical quantity, which is an element of the association structure D2, is greater than a threshold value for dependency selection, which is set for each statistical indicator. The fact that the statistical quantity is greater than the threshold value means that there is a dependency relationship between the sensors Xn. Note that the association structure graph output unit 360 can determine the dependency relationship between the sensors Xn from the association structure after the dependency relationship determination.

[0193] When the dependency between the sensors Xn is unidirectional, a single-sided arrow edge is displayed in the directed graph, and when the dependency between the sensors Xn is bidirectional, a double-sided arrow edge is displayed in the directed graph. For example, here, since there is a unidirectional dependency from the sensor X2 to the sensor X1, as shown in FIG. 27, a single-sided arrow edge (shown by D45G in FIG. 27) from the node indicating the sensor X2 (shown by D45B in FIG. 27) to the node indicating the sensor X1 (shown by D45A in FIG. 27) is displayed in the directed graph. Also, for example, here, there is a bidirectional dependency between the sensor X2 and the sensor X6, as shown in FIG. 27, a double-sided arrow edge (shown by D45H in FIG. 27) is displayed between the node indicating the sensor X2 and the node indicating the sensor X6 (shown by D45F in FIG. 27).

[0194] FIG. 31 is a flowchart for explaining an example of the operation of the abnormality factor estimating device 100 in a case where the abnormality factor estimating device 100 is configured to include the association structure graph output unit 360 in the first embodiment. The abnormality factor estimating device 100 performs operations as shown in the flowchart of Fig. 31 in addition to the operations described using the flowchart of Fig. 13. It is assumed that the operations as shown in the flowchart of Fig. 31 are performed at least once after the processes of steps ST1 to ST5 of Fig. 13 are performed. The operations as shown in the flowchart of Fig. 31 may be performed, for example, after the process of step ST5 of Fig. 13, before the process of step ST6 is performed, after the process of step ST6 is performed, or in parallel with the process of step ST6.

[0195] The associated structure graph output unit 360 receives an instruction to display an associated structure graph (step ST31). For example, the operator operates an input device such as a mouse or a keyboard to call up an input screen for an instruction to display an association structure graph on the display device 400. The operator inputs an instruction to display an association structure graph from the input screen for an instruction to display an association structure graph. The association structure graph output unit 360 accepts the instruction to display an association structure graph input by the operator. The associated structure graph output unit 360 outputs associated structure graph display information for displaying a graph related to the associated structure D2 to the display device 400 based on the associated structure D2, the abnormality detection sensor information D3, and the abnormality factor order estimation result D7 stored in the data storage unit 20 (step ST32). As a result, the display device 400 displays, for example, a graph screen as shown in FIG.

[0196] In this manner, by configuring the abnormality factor estimating device 100 to include the association structure graph output unit 360, the abnormality factor estimating device 100 can improve the explainability of information related to the estimation result of the abnormality factor.

[0197] <Modification> In the above-described first embodiment, the learning device 200 may be configured to include a learning sensor pair generation unit 370 that generates pairs of sensors 300 from among the multiple sensors 300 based on information (hereinafter referred to as "facility design information") that defines the connection relationships of multiple devices constituting the target facility and the multiple sensors 300 provided in the multiple devices. In this case, the associated structure learning unit 240 acquires learning sensor data for each pair of sensors 300 generated by the learning sensor pair generation unit 370, and learns the associated structure D2. For example, an operator or the like generates facility design information based on design drawings, and inputs the generated facility design information by operating an input device such as a mouse or keyboard, and the learning device 200 acquires the facility design information by accepting the input facility design information. FIG. 32 is a diagram showing an example of a configuration of the learning device 200 including a learning sensor pair generating unit 370 in the first embodiment. 32, for simplicity of explanation, the learning device 200 includes a learning sensor pair generation unit 370, an associated structure learning unit 240, and a learning data storage unit 220, as well as a learning sensor data acquisition unit 210 and a learning preprocessing unit 230. Although not shown in FIG. 32 for simplicity of explanation, the learning device 200 is connected to the abnormality factor estimation device 100.

[0198] The learning sensor pair generation unit 370 acquires equipment design information, and based on the acquired equipment design information, generates a pair of sensors 300 from among the multiple sensors 300 to be used when the associated structure learning unit 240 learns the associated structure D2. In detail, the learning sensor pair generating unit 370 determines a combination of two sensors 300 based on facility design information, and generates information (hereinafter referred to as "sensor pair information") D47 in which the combinations are listed. The learning sensor pair generating unit 370 outputs the generated sensor pair information D47 to the associated structure learning unit 240.

[0199] When sensor pair information D47 is output from the learning sensor pair generation unit 370, the associated structure learning unit 240 calculates statistics indicating the relationship between two different sensor data for multiple sensor data included in the learning data D23 based on the pair of sensors 300 set in the sensor pair information D47, and learns the associated structure D2 based on the calculated statistics.

[0200] The process in which the learning sensor pair generating unit 370 generates the sensor pair information D47 will be described below with a specific example using the drawings. FIG. 33 is a diagram for explaining the concept of an example of a method in which the learning sensor pair generation unit 370 generates sensor pair information D47 based on facility design information D46 when the learning device 200 is configured to include the learning sensor pair generation unit 370 in embodiment 1. Here, as an example, the number of sensors 300 is 8, and the sensors 300 are represented as sensors Xn (n=1 to 8). Also, here, as an example, it is assumed that the target facility is configured with five devices (device D46A, device D46B, device D46C, device D46D, and device D46E). Also, suppose that device D46A is provided with sensors X1 and X2, device D46B is provided with sensor X3, device D46C is provided with sensors X4 and X5, device D46D is provided with sensors X6 and X7, and device D46E is provided with sensor X8. Also, the device D46A is in a connection relationship with the device D46B, the device D46B is in a connection relationship with the device D46A, the device D46C, and the device D46D, the device D46C is in a connection relationship with the device D46B and the device D46E, the device D46D is in a connection relationship with the device D46B, and the device D46E is in a connection relationship with the device D46C. In addition, the devices in the connection relationship are connected by an undirected line in the design. In this case, the equipment design information D46 has contents as shown in FIG. 33. In addition, in FIG. 33, the equipment design information D46 is shown in a block diagram as an example, but this is only an example. The equipment design information D46 may be information that can understand the connection relationship of the multiple devices constituting the target equipment and the multiple sensors Xn provided in the multiple devices.

[0201] The learning sensor pair generating unit 370 generates a pair consisting of two different sensors Xn based on the facility design information D46. In detail, the learning sensor pair generating unit 370 generates a pair consisting of a sensor Xn attached to a certain device and a sensor Xn attached to a device connected to the device. 33, according to the facility design information D46, the devices D46A and D46B are in a connected relationship. In this case, the sensors X1 and X2 provided in the device D46A and the sensor X3 provided in the device D46B are also in a connected relationship. Therefore, the learning sensor pair generation unit 370 generates a pair of the sensor X1 and the sensor X3, and a pair of the sensor X2 and the sensor X3.

[0202] Furthermore, when two or more sensors Xn are provided in one device, the learning sensor pair generating section 370 generates pairs of different sensors Xn among the two or more sensors Xn provided in the same device. 33, according to the facility design information D46, the device D46A is provided with a sensor X1 and a sensor X2. Therefore, the learning sensor pair generating unit 370 generates a pair of the sensor X1 and the sensor X2.

[0203] In this way, the learning sensor pair generation unit 370 generates sensor pairs, for example, a pair of two sensors provided on different devices and connected to each other, and a pair of two different sensors provided on one device. It should be noted that the sensor pairs generated by the learning sensor pair generation unit 370 as described above are merely examples, and the learning sensor pair generation unit 370 may, for example, generate only pairs consisting of two sensors that are provided on different devices and that are connected to each other, or may generate only pairs consisting of two different sensors provided on one device.

[0204] As described above, when the learning device 200 is configured to include the learning sensor pair generation unit 370, in the operation of the learning device 200 described using the flowchart in Fig. 14, before the processing of step ST231 is performed, the learning sensor pair generation unit 370 generates sensor pair information D47 and outputs it to the associative structure learning unit 240, and in step ST231, the associative structure learning unit 240 acquires pairs of two different sensor data based on the learning data D23 and the sensor pair information D47. Note that the associative structure learning unit 240 sets all combinations of multiple sensor data included in the learning data D23 as pairs of sensor data based on the sensor pair information D47.

[0205] In this way, the learning device 200 is configured to include a learning sensor pair generation unit 370 that generates pairs of sensors 300 from among the multiple sensors 300 based on facility design information, and the associative structure learning unit 240 acquires learning sensor data for each pair of sensors 300 generated by the learning sensor pair generation unit 370 and learns the associative structure D2, so that the learning device 200 can reduce the possibility of detecting a dependency relationship between sensors 300 that are low in association due to their design, and can learn an associative structure D2 with improved reliability. As a result, the learning device 200 can provide the abnormality factor estimation device 100 with an associative structure D2 that can precisely estimate an abnormality factor.

[0206] As described above, the anomaly factor estimation device 100 according to the first embodiment is configured to include: a sensor data acquisition unit 10 that acquires a plurality of time-series sensor data collected by a plurality of sensors 300 provided in a plurality of equipment components that constitute a target equipment; an anomaly detection unit 30 that detects a plurality of anomaly detection sensors in which an anomaly has occurred among the plurality of sensors 300 based on the plurality of sensor data acquired by the sensor data acquisition unit 10; an anomaly detection order estimation unit 40 that estimates an anomaly detection order in which an anomaly has been detected for the plurality of anomaly detection sensors based on detection times at which the anomaly detection unit 30 detected the plurality of anomaly detection sensors; an anomaly propagation path tracking unit 50 that estimates an anomaly propagation order in which an anomaly has propagated based on anomaly detection sensor information D3 related to the plurality of anomaly detection sensors detected by the anomaly detection unit 30 and an estimation structure (association structure D2) in which dependency relationships between the equipment components are indicated; and an anomaly factor estimation unit 60 that estimates a cause of an anomaly based on the anomaly detection order estimated by the anomaly detection order estimation unit 40 and the anomaly propagation order estimated by the anomaly propagation path tracking unit 50. Therefore, the fault cause estimation device 100 can estimate the cause of an fault that has occurred in a facility, regardless of the complexity or scale of the facility.

[0207] The abnormality factor estimating device 100 is also configured to include an abnormality factor estimation result output unit 70 that outputs information related to the result of estimation of the abnormality factor by the abnormality factor estimating unit 60. Therefore, the anomaly cause estimation device 100 improves the interpretability and explainability of the estimation results of the anomaly cause to the operator. The anomaly cause estimation device 100 can reduce unnecessary inspection work by the operator and reduce the burden on the operator. Furthermore, the anomaly cause estimation device 100 can estimate the cause of anomaly using a quantitative index that does not rely on human subjectivity and present the basis for the estimation. The operator can determine the inspection sequence of the equipment with little effort.

[0208] Furthermore, the abnormality factor estimating device 100 can be configured to detect an abnormality detection sensor using a univariate type anomaly detection method. Therefore, the abnormality factor estimating device 100 can more appropriately detect an abnormality in which one piece of sensor data D1 changes independently.

[0209] Furthermore, the abnormality factor estimating device 100 can be configured to detect an abnormality detection sensor using a multivariate type abnormality detection method. Therefore, the abnormality factor estimating device 100 can more appropriately detect an abnormality in which the relationship between the multiple pieces of sensor data D1 changes.

[0210] Furthermore, the abnormality factor estimating device 100 can be configured to detect an abnormality using a univariate type anomaly detection method and a multivariate type anomaly detection method. Therefore, the abnormality factor estimating device 100 can more appropriately detect an abnormality in which one piece of sensor data D1 changes independently, or an abnormality in which the relationship between a plurality of pieces of sensor data D1 changes.

[0211] Furthermore, in the abnormality cause estimation device 100, the abnormality propagation path tracking unit 50 can be configured to estimate the abnormality propagation order based on the abnormality detection sensor information D3, equipment operation state information D31 indicating the operation state of the target equipment, and an estimation structure (association structure D32) indicating the dependency relationship between the equipment components according to the operation state of the target equipment. Therefore, the abnormality factor estimation device 100 can respond to changes in the dependency relationships between the sensors 300 that accompany changes in the operating state of the target equipment, and can precisely estimate the cause of the abnormality based on the association structure D32 with improved reliability.

[0212] In addition, the abnormality factor estimation device 100 can be provided with an association structure correction unit 330 that corrects the dependency relationship between sensor data for the estimated structure (association structure D2) based on dependency pair information D33 regarding a pair of sensors that have a dependency relationship among the multiple sensors 300 and dependency-free pair information D34 regarding a pair of sensors 300 that does not have a dependency relationship. Therefore, the abnormality factor estimation device 100 can improve the reliability of the estimated structure and perform precise estimation of the abnormality factor.

[0213] Furthermore, the abnormality factor estimation device 100 includes a relationship change estimation unit 340 that compares the estimated structure (associated structure D2), the estimated structure at the time of occurrence of an abnormality (associated structure D36), and the abnormality detection sensor information D3, and estimates a change in the relationship between the sensor data. The abnormality factor estimation unit 60 can be configured to estimate the cause of the abnormality by taking into consideration the change in the relationship between the sensor data estimated by the relationship change estimation unit 340, based on the abnormality detection order estimated by the abnormality detection order estimation unit 40 and the abnormality propagation order estimated by the abnormality propagation path tracking unit 50. Therefore, the fault factor estimation device 100 can improve the reliability of the fault factor order estimation result D7 and perform precise estimation of fault factors.

[0214] In addition, the abnormality factor estimation device 100 can be configured to include an abnormality factor equipment estimation unit 350 that estimates the cause of an abnormality on an equipment basis, based on equipment-attached sensor information D41 in which equipment installed in the target facility is associated with the sensors 300 installed in the equipment, the abnormality detection order estimated by the abnormality detection order estimation unit 40, and the abnormality propagation order estimated by the abnormality propagation path tracking unit 50. Therefore, the fault cause estimation device 100 enables the operator to efficiently identify the equipment causing the fault. Also, the fault cause estimation device 100 enables the operator to efficiently understand the order in which the equipment causing the fault should be inspected.

[0215] Furthermore, the abnormality factor estimation device 100 can be configured to include an associated structure graph output unit 360 that outputs associated structure graph display information for displaying a graph in which the estimated structure, the anomaly detection sensor, and the estimation result of the abnormality factor are associated, based on the estimated structure (associated structure D2), the abnormality detection sensor information D3, and information related to the estimation result of the abnormality factor estimated by the abnormality factor estimation unit 60. Therefore, the abnormality cause estimating device 100 can improve the explainability of information related to the estimation result of the abnormality cause.

[0216] As described above, the learning device 200 according to embodiment 1 is configured to include a learning sensor data acquisition unit 210 that acquires, as learning data candidates, multiple time-series sensor data collected by multiple sensors 300 installed in the target equipment during normal operation of the target equipment, and an associated structure learning unit 240 that uses the multiple learning data candidates acquired by the learning sensor data acquisition unit 210 as multiple learning data, calculates at least one statistic between the multiple learning data based on the learning data, and learns an estimated structure (associated structure D2) indicating the dependency relationships between the equipment components based on the calculated statistic. Therefore, the learning device 200 can comprehensively extract associations between the sensor data D1, and as a result, can provide an association structure D2 that suppresses overlooking of connection relationships between the sensors 300. By providing the anomaly cause estimation device 100 with an estimation structure (association structure D2) for tracking the sensor 300 that is the source of the anomaly, the learning device 200 can cause the anomaly cause estimation device 100 to more appropriately track the sensor 300 that is the source of the anomaly, thereby improving the accuracy of estimating the cause of the anomaly.

[0217] The learning device 200 also includes a learning preprocessing unit 230 that acquires multiple learning data to be used for learning based on multiple learning data candidates acquired by the learning sensor data acquisition unit 210, and the related structure learning unit 240 can be configured to calculate at least one statistic between the multiple learning data based on the learning data acquired by the learning preprocessing unit 230, and learn an estimated structure (related structure D2) based on the calculated statistic. Therefore, the learning device 200 can comprehensively extract associations between the sensor data D1, and as a result, can provide an association structure D2 that suppresses overlooking of connection relationships between the sensors 300. By providing the anomaly cause estimation device 100 with an estimation structure (association structure D2) for tracking the sensor 300 that is the source of the anomaly, the learning device 200 can cause the anomaly cause estimation device 100 to more appropriately track the sensor 300 that is the source of the anomaly, thereby improving the accuracy of estimating the cause of the anomaly.

[0218] Furthermore, in the learning device 200, the learning preprocessing unit 230 can be configured to select a plurality of learning data candidates whose variance values ​​are less than a selection threshold from among the plurality of learning data candidates acquired by the learning sensor data acquisition unit 210, and to acquire the selected plurality of learning data candidates as a plurality of learning data. Therefore, the learning device 200 can comprehensively extract associations between the sensor data D1, and as a result, can provide an association structure D2 that suppresses overlooking of connection relationships between the sensors 300. The learning device 200 provides the anomaly cause estimation device 100 with an estimation structure (association structure) for tracking the sensor 300 that is the source of the anomaly, thereby enabling the anomaly cause estimation device 100 to more appropriately track the sensor 300 that is the source of the anomaly, and thereby improving the accuracy of estimating the cause of the anomaly.

[0219] Furthermore, in the learning device 200, the relational structure learning section 240 can be configured to calculate statistics using waveform-based statistical indices. Therefore, the learning device 200 can track anomaly propagation based on the dependency of waveform similarity, and can provide an estimation structure (association structure D2) that can more appropriately estimate the cause of the anomaly.

[0220] Furthermore, the learning device 200 can be configured to calculate statistics using distribution-based statistical indices. Therefore, the learning device 200 can track anomaly propagation based on a dependency relationship that is distributionally similar, and can provide an estimation structure (association structure D2) that can more appropriately estimate anomaly causes.

[0221] Furthermore, the learning device 200 can be configured to calculate statistics using waveform-based statistical indices and distribution-based statistical indices. Therefore, the learning device 200 can track anomaly propagation based on a dependency relationship such as waveform or distribution similarity, and can provide an estimation structure (association structure D2) that can more appropriately estimate the cause of the anomaly.

[0222] The learning device 200 also includes a learning sensor pair generation unit 370 that generates pairs of sensors 300 from among the multiple sensors 300 based on equipment design information D46 that defines the connection relationships of the multiple devices that make up the target equipment and the multiple sensors 300 provided in the multiple devices, and the related structure learning unit 240 can be configured to acquire learning data based on the pairs of sensors 300 generated by the learning sensor pair generation unit 370 and learn an estimated structure (related structure D2). Therefore, the learning device 200 can reduce the possibility of detecting a dependency relationship between sensors 300 that is low in correlation by design, and can learn an association structure D2 with improved reliability. As a result, the learning device 200 can provide the abnormality factor estimation device 100 with an association structure D2 that can precisely estimate an abnormality factor.

[0223] In addition, in the present disclosure, any of the components of the embodiments may be modified or any of the components of the embodiments may be omitted. [Industrial Applicability]

[0224] The abnormality factor estimation device according to the present disclosure can estimate the cause of an abnormality that has occurred in a facility, regardless of the complexity or scale of the facility. [Explanation of symbols]

[0225] 1000 Precision diagnosis system, 100 Abnormality factor estimation device, 10, 310 Sensor data acquisition unit, 20, 320 Data storage unit, 30 Abnormality detection unit, 40 Abnormality detection order estimation unit, 50 Abnormality propagation path tracking unit, 60 Abnormality factor estimation unit, 70 Abnormality factor estimation result output unit, 330 Related structure correction unit, 340 Relationship change estimation unit, 350 Abnormality factor equipment estimation unit, 360 Related structure graph output unit, 200 Learning device, 210 Learning sensor data acquisition unit, 220 Learning data storage unit, 230 Learning preprocessing unit, 240 Related structure learning unit, 370 Learning sensor pair generation unit, 300 Sensor, 400 Display device, 1601 Processing circuit, 1602 Input interface device, 1603 Output interface device, 1604 Processor, 1605 Memory.

Claims

1. a sensor data acquisition unit that acquires a plurality of time-series sensor data collected by a plurality of sensors provided in a plurality of equipment components that constitute the target equipment; an anomaly detection unit that detects a plurality of anomaly detection sensors in which an anomaly has occurred among the plurality of sensors based on the plurality of sensor data acquired by the sensor data acquisition unit; an anomaly detection order estimation unit that estimates an anomaly detection order in which the anomaly was detected by the plurality of anomaly detection sensors based on detection times at which the anomaly detection unit detected the plurality of anomaly detection sensors; an anomaly propagation path tracing unit that estimates an anomaly propagation order in which the anomaly has propagated based on anomaly detection sensor information regarding the plurality of anomaly detection sensors detected by the anomaly detection unit and an estimated structure indicating a dependency relationship between the equipment components; an anomaly cause estimation unit that estimates a cause of the anomaly based on the anomaly detection order estimated by the anomaly detection order estimation unit and the anomaly propagation order estimated by the anomaly propagation path tracing unit; An abnormality cause estimation device comprising:

2. The predicted structure is represented by a matrix 2. The abnormality factor estimating device according to claim 1.

3. an abnormality factor estimation result output unit that outputs information about the result of the estimation of the abnormality factor by the abnormality factor estimation unit; 2. The abnormality factor estimating device according to claim 1, further comprising:

4. The anomaly detection unit detects the anomaly detection sensor using a univariate type anomaly detection method.

2. The abnormality factor estimating device according to claim 1.

5. The anomaly detection unit detects the anomaly detection sensor using a multivariate anomaly detection method.

2. The abnormality factor estimating device according to claim 1.

6. The anomaly detection unit detects the anomaly detection sensor using a univariate anomaly detection method and a multivariate anomaly detection method.

2. The abnormality factor estimating device according to claim 1.

7. The anomaly propagation path tracing unit estimates the anomaly propagation order based on the anomaly detection sensor information, equipment operation state information indicating an operation state of the target equipment, and the estimated structure indicating dependencies between the equipment components according to the operation state of the target equipment.

2. The abnormality factor estimating device according to claim 1.

8. a relational structure correcting unit that corrects the dependency relationship between the sensor data for the estimated structure based on dependency pair information regarding a pair of the sensors having a dependency relationship among the plurality of sensors and dependency non-dependency pair information regarding a pair of the sensors having no dependency relationship; 2. The abnormality factor estimating device according to claim 1, further comprising:

9. A relationship change estimation unit is provided which compares the estimated structure in a normal state with the estimated structure in an abnormal state based on the estimated structure in a normal state, the estimated structure in an abnormal state, and the abnormality detection sensor information, and estimates a change in the relationship between sensor data; The anomaly cause estimation unit estimates a cause of the anomaly based on the anomaly detection order estimated by the anomaly detection order estimation unit and the anomaly propagation order estimated by the anomaly propagation path tracing unit, while taking into consideration a change in a relationship between sensor data estimated by the relationship change estimation unit.

2. The abnormality factor estimating device according to claim 1.

10. an anomaly cause equipment estimation unit that estimates a cause of the anomaly for each of the equipment based on equipment-attached sensor information in which the equipment installed in the target facility is associated with the sensor installed in the equipment, the anomaly detection order estimated by the anomaly detection order estimation unit, and the anomaly propagation order estimated by the anomaly propagation path tracing unit; 2. The abnormality factor estimating device according to claim 1, further comprising:

11. an associated structure graph output unit that outputs associated structure graph display information for displaying a graph in which the estimated structure, the anomaly detection sensor, and the inference result of the cause of the abnormality are associated with each other, based on the estimated structure, the anomaly detection sensor information, and information on the inference result of the cause of the abnormality inferred by the anomaly factor inference unit.

2. The abnormality factor estimating device according to claim 1, further comprising:

12. A learning device that learns the estimation structure used to estimate the abnormality propagation order in the abnormality cause estimation device according to claim 1, comprising: a learning sensor data acquisition unit that acquires, as learning data candidates, a plurality of time series of the sensor data collected by a plurality of the sensors provided in a plurality of equipment components constituting the target equipment during normal operation of the target equipment; a related structure learning unit that calculates at least one statistic between the plurality of learning data based on the plurality of learning data candidates acquired by the learning sensor data acquisition unit as a plurality of learning data, and learns the estimated structure in which the dependency relationships between the equipment components are indicated based on the calculated statistic. A learning device comprising:

13. a learning preprocessing unit that acquires a plurality of pieces of learning data to be used for learning based on the plurality of learning data candidates acquired by the learning sensor data acquisition unit; The related structure learning unit calculates at least one statistic between a plurality of the learning data based on the learning data acquired by the learning preprocessing unit, and learns the estimated structure based on the calculated statistic.

13. The learning device according to claim 12.

14. The learning preprocessing unit selects a plurality of the learning data candidates having a variance value less than a selection threshold from the plurality of the learning data candidates acquired by the learning sensor data acquisition unit, and acquires the selected plurality of the learning data candidates as a plurality of the learning data.

14. The learning device according to claim 13.

15. The related structure learning unit calculates the statistics using a waveform-based statistical index.

13. The learning device according to claim 12.

16. The association structure learning unit calculates the statistics using a distribution-based statistical index.

13. The learning device according to claim 12.

17. The related structure learning unit calculates the statistics using a waveform-based statistical index and a distribution-based statistical index.

13. The learning device according to claim 12.

18. a learning sensor pair generation unit that generates pairs of sensors from among the plurality of sensors based on facility design information that defines a connection relationship between a plurality of devices constituting the target facility and a plurality of sensors provided in the plurality of devices; The related structure learning unit acquires the learning data based on the sensor pairs generated by the learning sensor pair generation unit, and learns the estimated structure.

13. The learning device according to claim 12.

19. An abnormality factor estimating device according to any one of claims 1 to 11, A learning device according to any one of claims 12 to 18. A precision diagnostic system equipped with

20. A step in which a sensor data acquisition unit acquires a plurality of time-series sensor data collected by a plurality of sensors provided in a plurality of equipment components constituting a target equipment; An abnormality detection unit detects a plurality of abnormality detection sensors in which an abnormality is occurring among the plurality of sensors based on the plurality of sensor data acquired by the sensor data acquisition unit; an anomaly detection order estimation unit estimating an anomaly detection order in which the anomaly detection unit detected the occurrence of the anomaly for the plurality of anomaly detection sensors based on detection times at which the anomaly detection unit detected the plurality of anomaly detection sensors; an anomaly propagation path tracing unit estimating an anomaly propagation order in which the anomaly has propagated based on anomaly detection sensor information regarding the plurality of anomaly detection sensors detected by the anomaly detection unit and an estimated structure indicating a dependency relationship between the equipment components; an anomaly cause estimation unit estimating a cause of the anomaly based on the anomaly detection order estimated by the anomaly detection order estimation unit and the anomaly propagation order estimated by the anomaly propagation path tracing unit; The abnormality cause estimation method includes the steps of:

Images