WO2017109903A1

WO2017109903A1 - Malfunction cause estimation device and malfunction cause estimation method

Info

Publication number: WO2017109903A1
Application number: PCT/JP2015/086085
Authority: WO
Inventors: 康平丸地
Original assignee: 株式会社東芝
Priority date: 2015-12-24
Filing date: 2015-12-24
Publication date: 2017-06-29
Also published as: JPWO2017109903A1

Abstract

A problem to be addressed by the present invention is to allow a more reliable diagnosis by combining cause estimation models when carrying out a cause estimation. Provided is a malfunction cause estimation device, comprising: a data acquisition unit which acquires sensor data of a sensor which has been installed in a facility; a first estimation unit which, on the basis of the sensor data, estimates a cause of a malfunction of the facility from a first estimation model; and a correlation table storage unit which stores correlation tables which provide correlations among estimated causes and a second estimation model which augments the estimated causes. The malfunction cause estimation device further comprises: a second estimation unit which, if an association has been made with the second estimation model via the correlation tables, additionally estimates the cause of the malfunction on the basis of the second estimation model; and a display unit which displays the estimated cause.

Description

Abnormal cause estimation apparatus and abnormality cause estimation method

Embodiments of the present invention relate to an apparatus and a method for estimating a cause of an abnormality using data measured by a sensor.

∙ Devices and systems that acquire data from sensors and perform desired control perform desired operations while performing self-diagnosis to ensure that they are always operating in the correct state for stable operation. When an abnormal state is detected, a notification to that effect is sent to inform the operator or user that an abnormality has occurred. The worker or user who received the notification identifies the cause of the abnormal state and takes appropriate measures according to the cause.

The abnormal state detected by the system when an error is issued is the abnormal state at the time of detection, and the cause of the abnormality is unknown. In a system in which a large number of parts are combined and operated, the parts operate while affecting each other, and therefore, the part that has issued an abnormality does not necessarily have a cause. There may be a failure of an adjacent component or a failure of another component which is not adjacent but repeatedly propagates the influence of the failure and is distant from the structure.

Only after the cause of the failure is known, how to recover is clarified. Accurate and quick estimation of the cause leads to a reduction in man-hours for these recovery operations.

As described above, there are a method using a diagnosis rule, a method using machine learning, and a method using a physics or chemistry model as a system that supports accurate and quick cause investigation after detecting an abnormality or failure.

The method using the diagnosis rule is a method of making a diagnosis based on rules of expert knowledge and empirical rules such as “cause C when the value of thermometer A is greater than B”. The method using machine learning is a method of making a diagnosis by constructing a diagnosis model obtained by machine learning of past data for which cause investigation has been completed and classifying which case is similar to the past case. The method using a physical or chemical model is a method of simulating system behavior using, for example, a physical law or chemical formula, and detecting a difference between a simulation result and a measured value to identify an abnormality and identify a cause.

課題 A common issue for these diagnostic models is to increase the accuracy (accuracy) of the diagnosis. Depending on the diagnostic model, there are cases where a plurality of causes are listed as candidates, and erroneous determination is likely to occur for a specific failure, and it is necessary to construct a model with higher estimation accuracy.

JP-A-8-202444 JP 2009-53938 A

The problem to be solved by the present invention is to enable diagnosis with higher accuracy by performing cause estimation by combining cause estimation models.

An abnormality cause estimation apparatus according to an embodiment is an abnormality cause estimation apparatus that estimates an abnormality cause of the equipment based on sensor data of a sensor installed in the equipment, and includes a data acquisition unit that acquires the sensor data, and the sensor A first storage unit that stores a first estimation model that estimates an abnormality cause of the facility based on data, a first estimation unit that obtains a first estimation cause based on the first estimation model, and A second storage unit that stores a second estimation model that supplements the first estimation cause, and a correspondence table storage unit that stores a correspondence table that associates the first estimation cause with the second estimation model And when the first estimation cause and the second estimation model are associated with each other by the correspondence table, a second estimation unit that obtains a second estimation cause based on the second estimation model; The first probable cause and the second And a display unit that displays the probable cause.

In addition, the abnormality cause estimation method of the embodiment includes a data acquisition unit that acquires sensor data of a sensor installed in equipment, and a first estimation model that estimates the cause of abnormality of the equipment. A first estimation unit, a correspondence table storing a correspondence table associating the first estimated cause with a second estimated model that supplements the first estimated cause, and a second estimated cause based on the second estimated model An abnormality cause estimation method in an abnormality cause estimation device comprising: the sensor data is acquired by the data acquisition unit; and the first estimation unit is based on the sensor data. The first estimation model estimates the first estimation cause, and in the second estimation unit, the first estimation cause and the first estimation based on the correspondence table stored in the correspondence table storage unit Two estimation models If that is correlated, the second estimation model is a method for estimating a second probable cause.

It is a block diagram of the abnormality cause estimation apparatus of 1st Embodiment. It is a flowchart of an abnormality cause estimation process. It is a block diagram of a 1st estimation part and a 1st estimation model. It is an example of arrangement of presumed causes. It is an example of a rule-based diagnostic logic. It is an example of a correspondence table. It is a block diagram of a 2nd estimation part and a 2nd estimation model. It is an example of a display by a display part. It is an example of a mixing matrix. It is a flowchart of the abnormality cause estimation process of a 1st modification. It is a model selection screen image figure to a user. It is a flowchart of the abnormality cause estimation process of a 2nd modification. It is a block diagram of the abnormality cause estimation apparatus of 2nd Embodiment. It is a flowchart of the 2nd estimation model construction process for multiple causes. It is an example of a diagnostic result of a model evaluation part. It is an example of a mixing matrix created by the model evaluation unit. It is a flowchart of preparation of the 2nd estimation model for multiple causes. It is a flowchart of the 2nd estimation model construction process for the result which is easy to make a mistake. It is Example 1 of the mixing matrix with the result which is easy to make a mistake. It is Example 2 of the mixing matrix with the result which is easy to make a mistake. It is Example 3 of the mixing matrix with the result which is easy to make a mistake. It is a block diagram of the abnormality cause estimation apparatus of 3rd Embodiment. It is a flowchart which builds the 2nd estimation model designated from the outside. It is a block diagram of the abnormality cause estimation apparatus of 4th Embodiment. It is a flowchart which builds the 1st estimation model designated from the outside.

(First embodiment)
FIG. 1 is a block diagram showing an abnormality cause estimation apparatus of the first embodiment, and FIG. 2 is a flowchart thereof.

The abnormality cause estimation apparatus according to the first embodiment includes a data acquisition unit 20 that acquires sensor data 10 of equipment to be diagnosed, a first estimation unit 30 that performs cause estimation based on a first estimation model 40A, a first A first storage unit 40 that stores the estimation model 40A, a second estimation unit 50 that performs detailed cause estimation based on the second estimation model 60A, a second storage unit 60 that stores the second estimation model 60A, A correspondence table storage unit 70 that stores the correspondence table 70A and a display unit 80 are provided.

Sensor data 10 is sensor data measured by a large number of sensors arranged at various locations of the diagnosis target equipment, and is time-series data composed of measured values and measurement times for each sensor.

Sensor data 10 may include state variables inside the system. The data acquisition unit 20 includes a communication unit, and acquires a measurement value of a sensor installed in the facility constantly or at a constant timing. The equipment is connected using a USB or a connection port. In addition, log data accumulated in a certain amount of memory in the facility may be acquired via a storage medium such as an SD card or a USB memory (S201).

The first estimation unit 30 estimates the cause of the abnormality using the first estimation model 40A in the first storage unit 40 with the sensor data 10 of the facility as an input (S202). Specifically, the first estimation unit 30 corresponds to a calculation location for performing cause estimation, the first storage unit 40 corresponds to a storage device such as a hard disk, and the first estimation model 40A performs cause estimation. This corresponds to a program using an algorithm. The cause estimation in the first estimation unit 30 is executed by an arithmetic device such as a CPU (Central Processing Unit).

FIG. 3 shows the configuration of the first estimation unit 30 and the first estimation model 40A. The first estimation model 40A used in the present embodiment includes an input feature quantity list 41, an estimation logic 42, and model meta information 43. The first estimating unit includes a feature calculating unit 31, a model executing unit 32, and a result organizing unit 33. The feature calculation unit 31 calculates a feature amount necessary for the estimation logic 42 from the sensor data 10. The feature quantity required by the estimation logic 42 is defined in the input feature quantity list 41.

For example, the input feature quantity list 41 has a description expressing characteristics such as an average value of the measurement data of the sensor A in the input feature quantity list 41, and the feature calculation unit 31 interprets it and calculates a desired feature quantity. Note that the input feature quantity list 41 can be omitted by determining the feature quantity to be used in advance.

The model execution unit 32 performs cause estimation based on the estimation logic 42 using the feature amount obtained by the feature calculation unit 31. The estimation logic 42 performs cause estimation from the input feature quantity. The cause estimation here is to calculate a value (here called accuracy) that quantifies the possibility of each cause. The result organizing unit 33 finally organizes the accuracy of each cause. FIG. 4 shows an example. When the identification is uniquely specified, the accuracy of the identified cause is 1, and when the accuracy is limited to a plurality, the accuracy of the cause is 1. When the accuracy is originally calculated, the value is used as it is. In addition, all possible causes may be listed in the cause name column, but only main ones may be listed as shown in FIG. 4 and the others may be arranged in other ways.

The estimation logic 42 is based on logic by machine learning. The logic by machine learning is a logic constructed using an algorithm that solves a classification problem. Examples of these algorithms include decision trees, random forests, SVM (Support Vector Machine), and neural networks. An algorithm combining these algorithms may be used.

The model meta information 43 is meta information related to the model. Similar to the input feature 41, it is not always necessary, but if this information is present, the display unit 80 can provide a user-friendly display. Examples of the model meta information 43 include a model name, a model correct answer rate, a model mixing matrix, and a cause name list to be estimated by the model.

The first estimation model 40A is preferably an algorithm that can cover all estimated causes.

Returning to the flowchart of FIG. 2, in S203, the second estimation unit 50 determines whether or not the second estimation model 60A corresponding to the cause of the abnormality cause obtained in S202 exists in the second storage unit 60. Is performed by examining the correspondence table 70A in the correspondence table storage unit 70.

Similar to the first estimation unit 30, the second estimation unit 50 corresponds to a calculation location, the second storage unit 60 and the correspondence table storage unit 70 correspond to a storage device such as a hard disk, and the second estimation model is This corresponds to a program using an algorithm for estimating the cause. The cause estimation in the second estimation unit 50 is executed using an arithmetic device such as a CPU. Moreover, the same storage device may be used as the storage device corresponding to the first storage unit 40, the second storage unit 60, and the correspondence table storage unit 70.

The correspondence table 70A and the second estimation model 60A are created in advance from past data and empirical rules. Specifically, when the estimation result from the past data by the first estimation model 40A includes a plurality of estimation causes, or the estimation cause is different from the actual estimation cause (probable estimation cause 2), the second estimation model 60A that supplements the estimation cause is constructed. The correspondence table 70A is a table in which the correspondence between the estimation cause in the first estimation unit 30 and the constructed second estimation model 60A is created as a table. That is, the second estimation model 60A is a model used to increase the accuracy of the estimation cause of the first estimation unit 30.

Since the second estimation model 60A to be selected differs depending on the estimation cause of the first estimation unit 30, the second estimation model 60A to be additionally estimated is associated using the correspondence table 70A.

By using the correspondence table 70A, the second estimation model 60A can be easily extracted.

FIG. 6 is an example of the correspondence table 70A. The estimation cause is associated with the second estimation model 60A. Further, it is possible to associate with the second estimation model 60A using a conditional expression based on the accuracy of the estimation cause.

In S203, the estimated cause of the first estimating unit 30 is compared with the correspondence table 70A, and if there is a second estimated model corresponding to the estimated cause (in the case of Yes), the process proceeds to S204, and if not, the process proceeds to S205. Proceed to

In S204, the second estimation unit 50 performs detailed cause estimation using the second estimation model 60A obtained from the correspondence table 70A in the correspondence table storage unit 70. FIG. 7 is a configuration diagram of the second estimation unit 50 and the second estimation model 60A. Comparing FIG. 3 and FIG. 7, a second estimation model selection unit 51 is added to the second estimation unit 50, and the other features calculation unit 52 and model execution unit 53 are the same as those of the first estimation unit 30. And a result organizing unit 54. The second estimation unit 50 is a part that determines the second estimation model 60A used from the correspondence table 70A. If the second estimation model is determined, the remaining operations are estimated in the same manner as the first estimation unit 30 and thus will not be described.

When S203 is No or after S204 is completed, the estimated cause of the first estimating unit 30 or the estimated causes of both the first estimating unit 30 and the second estimating unit 50 are displayed on the display unit 80 in S205. indicate. That is, when S203 is No, the display unit 80 displays the estimation cause and the estimation accuracy of the first estimation model so that the user can easily understand. When S203 is Yes, the display unit 80 can display not only the estimated cause of the first estimated model but also the detailed estimated cause based on the second estimated model. The display unit 80 corresponds to a computer monitor, a liquid crystal monitor of a portable terminal, or the like.

FIG. 8 shows an example of a screen display in which the estimated causes based on the first estimated model and the estimated causes of the second estimated model are arranged. On the left side of FIG. 8 is an example in which bar graphs are displayed in ascending order of diagnostic accuracy of the first estimation model 40A, and the recall of each cause A and B of the first estimation model 40A is represented by a line graph. . In order to display the recall rate, it is assumed that the model meta information 43 has a mixing matrix. The mixing matrix is a table in which the diagnosis results of the diagnosis logic and the actual results are arranged (FIG. 9). In the example of FIG. 9, the number of times that cause A is correctly diagnosed as cause A is 10 times, and the number of times that cause C is mistakenly diagnosed as cause B is one. The recall is an index representing the certainty of the diagnosis result of the diagnosis logic, and is a ratio that the estimated cause is correct with respect to a specific estimated cause. In the example of FIG. 9, the recall of cause A is 10/17, the recall of cause B is 5/9, and the recall of cause C is 1/3.

On the right side of FIG. 8, a more detailed diagnosis result based on the second estimation model is represented by a pie chart, and the model AB, which is the name of the second estimation model, and the diagnostic accuracy rate thereof are displayed.

The algorithm related to the estimation logic 42 of the abnormality cause estimation device of the embodiment is based on logic based on machine learning, but other than that, rule-based logic and logic based on a physical or chemical model are also conceivable.

FIG. 5 shows an example of logic based on the rule base and an example based on the if-then rule. When the cause of any of the causes A, B, and C is estimated, a threshold value is provided for the feature amount calculated by the feature calculation unit 31 and divided into cases. In this case, if the feature amount A is 90, the feature amount B is 70, and the feature amount C is 100, the cause B or the cause C is estimated, the feature amount A is 90, the feature amount B is 40, and the feature amount C is 100. If there is any, it is estimated as other. These threshold values are determined based on past data and empirical rules.

Logic based on a physical or chemical model is a method of estimating the cause of an abnormality by simulating the system behavior to be compared and looking at the difference between the measured value and the simulated value. As described above, in addition to machine learning, various logics can be used to obtain an estimated cause.

As described above, according to the first embodiment, since the estimation by the second estimation model 60A that supplements the cause of the estimation by the first estimation model 40A is also performed, the first estimation model 40A alone is used for estimation. Makes it possible to estimate the cause with high accuracy. In addition, the correspondence table 70A can easily extract the second estimation model 60A that improves the accuracy of the estimation cause in the first estimation unit 30.

(First modification of the first embodiment)
FIG. 10 is a first modification of the flowchart of the cause estimation process. Since step S1001 is different from FIG. 2, this portion will be described. Process S1001 is a process of confirming with the user whether to perform estimation using the second estimation model when a correspondence table exists. If the user wishes to estimate, the cause is estimated using the second estimation model (S204). When the user does not wish to estimate (for example, when there is no estimation instruction), the estimation cause is displayed without estimating the second estimation model (S205).

FIG. 11 shows an example of a screen display when asking the user whether to use the estimated model. The display of the first estimation model is the same as in FIG. At this time, the accuracy and the reproducibility of cause A and cause B are as high as each other. It is assumed that model AB and model AB + are listed as candidates from the correspondence table 70A. Model AB is a model for diagnosing either cause A or cause B, and model AB + is a model for diagnosing whether cause A and cause B are occurring simultaneously. This information can be obtained by the user by looking at the diagnosis target (FIG. 11). The diagnosis target can be output by registering this information in the model meta information 63. Similarly, by registering the correct answer rate in the model meta information 63, it can be presented to the user as shown in FIG. The user examines whether to estimate by referring to these pieces of information, and if so, marks the selection checklist and informs the apparatus of the model to be estimated by pressing the estimation execution button based on the selected model.

As shown in FIG. 11, by allowing the user to browse the second estimated model list, it is possible to execute a model other than the second estimated model selected by the system.

As described above, in the first modification of the first embodiment, since the user can select which estimation model to use when performing the cause estimation in the second estimation unit 50, it is used in detail for diagnosis of the cause of the abnormality. In the case of a person, an estimated cause with higher accuracy can be obtained.

(Second modification of the first embodiment)
FIG. 12 is a second modification of the flowchart of the cause estimation process. In FIG. 2, after the estimation with the first estimation model 40A, the estimation with the second estimation model 60A is performed only once, but in FIG. 12, the second estimation model 60A with respect to the second estimation model 60A is performed. We also estimate.

This is because, in the second estimation unit 50, for example, when the accuracy of ABC is high, there is a possibility that the result cannot be specified by one cause estimation. In that case, the estimation cause in the second estimation unit 50 is associated with the second estimation model by the correspondence table 70A, and the second estimation model is estimated by the second estimation model. This cause estimation is repeated until the estimation cause in the second estimation unit 50 and the second estimation model are not associated in the correspondence table 70A. For this purpose, it is also necessary to define a correspondence table 70A for each second estimation model 60A.

As described above, in the second modification of the first embodiment, when there are a plurality of estimation causes estimated by the second estimation unit 50, the estimation cause and the second estimation model of the second estimation unit 50 are further added. Can be uniquely determined by associating them with the correspondence table 70A.

(Second Embodiment)
FIG. 13 is a block diagram showing the second embodiment. Implementation procedures of this embodiment include a procedure for performing cause estimation and a procedure for constructing a second estimation model. Since the former is not different from the first embodiment, a procedure for constructing a second estimation model newly increased in the present embodiment will be described. FIG. 14 is a flowchart of this procedure, and the components of FIG. 13 will be described according to the procedure.

First, the first evaluation model 40A is evaluated by the model evaluation unit 90 using the sensor abnormality data in the sensor abnormality data storage unit 100 (S1401). The sensor abnormality data storage unit corresponds to a database storing sensor abnormality data, and is stored in a hard disk, a USB memory, a ROM, or the like. Further, the sensor abnormality data storage unit may be in an external server or the like, and sensor abnormality data may be acquired therefrom. The model evaluation unit 90 is a calculation location and is processed by a CPU or the like.

The sensor abnormality data is data in which the cause of the abnormality is added to the sensor data 10 when an abnormality of the facility has occurred in the past. The model evaluation unit 90 uses the data in the sensor abnormality data storage unit 100 to evaluate the first estimation model 40A. The evaluation procedure is the same as that in S202 of FIG. 2, and the diagnosis result and the mixing matrix of each data are calculated as the evaluation result. The diagnosis result is obtained by calculating the accuracy of each cause for each sensor abnormality data, and can be organized as shown in FIG.

The mixing matrix is an arrangement of the number of diagnosis results and actual results, and is arranged as shown in FIG. The diagnosis result is calculated based on the accuracy shown in FIG. The cause with the highest accuracy may be selected, or a threshold value may be provided for each cause, and all the causes that are equal to or higher than the threshold value may be selected. You may select using the ratio and difference of a threshold and accuracy. When a plurality is selected, a mixing matrix including a column such as “cause A or cause C” is obtained as shown in FIG. If the cause of abnormality is a combination of a plurality of causes, a mixed matrix including rows such as “cause A and cause C” is obtained as shown in FIG.

Subsequently, the process proceeds to a second estimation model creation process (S1402) for multiple causes. Here, when there are a plurality of estimation causes of the first estimation model 40A, a second estimation model that performs estimation narrowed down from a plurality of candidates is created. FIG. 17 shows the detailed procedure. First, it is confirmed whether the case which estimates multiple causes is high frequency (S1701). This can be confirmed from the mixing matrix.

In the example of FIG. 16, it may be estimated that “cause A or cause C”, and the number of occurrences is the sum of the columns. Whether the number of occurrences is high can be determined by setting a threshold value. As the threshold value, an absolute number may be set or a ratio of the number of sensor abnormality data may be set.

If the frequency is not high (No in S1701), the process ends. When the frequency is high (Yes in S1701), the model construction unit 110 acquires sensor abnormality data of multiple causes from the sensor abnormality data storage unit 100 (S1702). In the case of FIG. 16, the sensor abnormality data of the cause A and the sensor abnormality data of the cause C are acquired from the sensor abnormality data storage unit 100.

Subsequently, the model construction unit 110 constructs a model for classifying a plurality of causes using the acquired sensor abnormality data (S1703). In the case of FIG. 16, a model is constructed as a classification problem of cause A sensor abnormality data and cause C sensor abnormality data. The model is constructed using general machine learning. Major algorithms include decision trees, random forests, SVMs, neural networks, and the like.

In the second estimation model 60A of the present embodiment, the second estimation model 60A is constructed for the estimation causes that are difficult to discriminate among the estimation causes in the first estimation model 40A. The configuration may be different from the second estimation model 60A.

Subsequently, the model construction unit evaluates the constructed second estimated model (S1704). The evaluation is performed using the sensor abnormality data in the sensor abnormality data storage unit. Since the cause of the abnormality is given to the sensor abnormality data, the accuracy of the model can be understood by evaluating the constructed second estimated model with the sensor abnormality data. The accuracy is calculated mainly using modeling fitting errors and cross validation. At this time, a plurality of classification algorithms can be employed to select a model having the highest evaluation result. The model construction unit 110 is a program that uses an algorithm, and is also a calculation part that evaluates the accuracy of the constructed model. These calculations are performed using a CPU or the like.

Subsequently, it is determined whether or not the second estimation model is to be adopted by judging whether the evaluation result (S1705) is good or bad. Whether it is adopted is determined by the accuracy rate and the threshold value of the F value. Based on the evaluation result of the first estimation model in S1401 of FIG. 14, a threshold value with higher accuracy than the first estimation model is set. If the evaluation result is bad (poor in S1705), the process ends as it is. If the evaluation result is good (good in S1705), the model updating unit 120 updates the second estimated model 60A in the second storage unit 60 and the correspondence table 70A in the correspondence table storage unit 70 (S1706).

In the case of the process of S1702, in the case of FIG. 16, it has been described that the actual cause is the sensor abnormality data of the cause A and the sensor abnormality data of the cause C are acquired from the sensor abnormality data storage unit 100. Sensor abnormality data in which cause C occurs at the same time may be acquired, and a model for classifying the three cases in S1703 may be constructed.

When there are a plurality of combinations of causes having a high frequency of occurrence of multiple causes in S1701, the processes of S1702 to S1705 may be repeated for the number of combinations.

Subsequently, the process proceeds to the second estimation model creation process (S1403 in FIG. 14) for results that are likely to be mistaken. Here, when there is a case where the estimation cause of the first estimation model 40A is easy to be mistaken, a second estimation model for confirming whether an erroneous estimation is performed is created. FIG. 18 shows the detailed procedure. First, it is confirmed whether there are frequent cases of wrong cause estimation (S1801). This can be confirmed from the mixing matrix. In the example of FIG. 19, the cause B is estimated, but the actual cause is A, and there are many erroneous estimates. Whether the number of occurrences is high can be determined by setting a threshold value. As the threshold value, an absolute number may be set or a ratio of the number of sensor abnormality data may be set.

If the frequency is not high (No in S1801), the process is terminated. If the frequency is high (Yes in S1801), the model construction unit 110 acquires sensor abnormality data that is easily mistaken from the sensor abnormality data storage unit 100 (S1802). Data to be acquired is determined according to cases that are easy to make mistakes. In the case of FIG. 19, there are many cases where the case of cause A is mistaken as cause B, but there are few cases where the case of cause B is mistaken as cause A. For this reason, it can be determined that it is difficult to distinguish between cause A and cause B in the case where the first estimation model 40A estimates cause B. Therefore, in S1802, data in which the actual cause is cause A and cause B among the cases estimated by the first estimation model 40A as cause B is acquired, and a model for classifying these is created in S1803.

In the case of FIG. 20, there are many cases where the case of cause A is mistaken as cause B, and the case of cause B is mistaken as cause A. Therefore, it can be determined that the classification of cause A and cause B is a difficult problem. Therefore, in S1802, data whose actual causes are cause A and cause B are acquired, and a model for classifying these is created in S1803.

In the case of FIG. 21, there are many cases where the case of cause A is mistaken as cause B, and the case of cause C is mistaken as cause B. Since the opposite is not true, it can be judged that it is difficult to distinguish between cause A, cause B, and cause C in the case estimated as cause B. Therefore, in S1802, data in which the actual cause is cause A, cause B, and cause C among the cases estimated by cause 40A of the first estimation model 40A are acquired, and a model for classifying these is created in S1803. .

Since S1803 to S1806, which are subsequent processes, are the same as S1703 to S1706 in FIG. 17, description thereof will be omitted.

In FIG. 14, the procedure for performing both the second estimated model creation process for multiple causes (S1402) and the second estimated model creation process for erroneous results (S1403) has been introduced, but only one of them is performed. It doesn't matter.

As described above, according to the second embodiment, when the estimated cause of the first estimated model 40A causes a plurality of estimated causes, or when the estimated cause of the first estimated model 40A is likely to be erroneous. By automatically constructing a model with high estimation accuracy based on past sensor abnormality data and performing estimation again, it is possible to estimate the cause with higher accuracy.

(Third embodiment)
FIG. 22 is a block diagram showing the third embodiment. The implementation procedure includes a procedure for performing cause estimation, a procedure for constructing the second estimation model 60A, and a procedure for constructing a model designated from the outside. The cause estimation procedure is the same as that of the first embodiment and the second estimation model 60A is the same as the second embodiment. Therefore, a newly specified externally specified model is established in the third embodiment. The procedure to do is explained. The procedure is the flowchart of FIG.

First, the external request acquisition unit 130 acquires an external request. Here, the external request is the specification of the model to be created, and is information about the model learning data and the model algorithm. The specifications of the model learning data include the actual cause and the period during which the data was obtained. If the request from the outside is acquired, the model construction unit 110 acquires sensor abnormality data designated from the outside from the sensor abnormality data storage unit 100 (S2302). Subsequently, the model construction unit 110 constructs a second estimated model 60A using the acquired sensor abnormality data and an externally designated algorithm (S2303).

Next, the second estimation model 60A is evaluated (S2304), and if the evaluation result is good, the model update unit 120 updates the correspondence table 70A in the second storage unit 60 and the correspondence table storage unit 70. (S2306). Since this procedure is the same as S1704-1706, description thereof is omitted.

As described above, according to the third embodiment, it is possible to construct the second estimation model 60A designated from the outside. If the user is familiar with the diagnosis of the cause of the abnormality and has already narrowed down the cause of the abnormality, the second estimation model 60A can be constructed from the narrowed down cause. Therefore, it is possible to perform estimation intended by the user, leading to improvement in estimation accuracy.

(Fourth embodiment)
FIG. 24 is a block diagram showing the fourth embodiment. In the third embodiment, the second estimation model 60A designated from the outside can be constructed, whereas in the present embodiment, the first estimation model 40A designated from the outside can be constructed.

Here, the external request acquisition unit 130 acquires an external request, and the model construction unit 110 acquires externally designated sensor abnormality data from the sensor abnormality data storage unit 100 (S2502). The model construction unit 110 constructs the first estimated model 40A using the acquired sensor abnormality data and an externally designated algorithm (S2503). Next, the first estimation model 40A constructed by the model construction unit 110 is evaluated (S2504), and if the evaluation result is good, the model storage unit 120 updates the first storage unit 40 (S2506).

In order to change the first estimation model 40A, it is preferable to redefine the correspondence table 70A and the second estimation model 60A.

As described above, according to the fourth embodiment, it is possible to construct the first estimation model 40A designated from the outside. When it is necessary to reconstruct the first estimation model 40A, such as when a new abnormality cause occurs, or when it is necessary to redefine the first estimation model 40A and perform more accurate estimation from the outside. By designating, the first estimation model 40A can be constructed.

Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 10 Sensor data 20 Data acquisition part 30 1st estimation part 31 Feature calculation part 32 Model execution part 33 Result arrangement | positioning part 40 1st memory | storage part 40A 1st estimation model 41 Input feature-value list | wrist 42 Estimation logic 43 Model meta information 50 Second estimation unit 51 Second estimation model selection unit 52 Feature calculation unit 53 Model execution unit 54 Result organizing unit 60 Second storage unit 60A Second estimation model 61 Input feature list 62 Estimation logic 63 Model meta information 70 Correspondence table storage unit 70A Correspondence table 80 display unit 90 model evaluation unit 100 sensor abnormality data storage unit 110 model construction unit 120 model update unit 130 user request acquisition unit

Claims

An abnormality cause estimation device for estimating an abnormality cause of the equipment based on sensor data of a sensor installed in the equipment,
A data acquisition unit for acquiring the sensor data;
A first storage unit for storing a first estimation model for estimating an abnormality cause of the equipment based on the sensor data;
A first estimation unit for obtaining a first estimation cause based on the first estimation model;
A second storage unit that stores a second estimation model that supplements the first estimation cause;
A correspondence table storage unit that stores a correspondence table that associates the first estimated cause with the second estimated model;
A second estimation unit that obtains a second estimated cause based on the second estimated model when the first estimated cause and the second estimated model are associated with each other by the correspondence table;
A display unit for displaying the first estimated cause and the second estimated cause;
An abnormality cause estimation device comprising:
A sensor abnormality data storage unit that accumulates sensor abnormality data in which the cause of abnormality of the facility is given to sensor data when the facility has an abnormality;
A model evaluation unit for obtaining an evaluation result of the first estimation model based on the sensor abnormality data;
A model construction unit for constructing a second estimation model based on the evaluation result of the model evaluation unit;
In accordance with the constructed second estimation model, a model update unit that updates the second storage unit and the correspondence table storage unit;
The abnormality cause estimation device according to claim 1, comprising:
A sensor abnormality data storage unit that accumulates sensor abnormality data in which the cause of abnormality of the facility is given to sensor data when the facility has an abnormality;
A model evaluation unit for obtaining an evaluation result of the first estimation model based on the sensor abnormality data;
An external request acquisition unit for acquiring external requests;
A model construction unit that constructs a second estimation model based on the evaluation result of the model evaluation unit and the data of the external request acquisition unit;
The abnormality cause estimation device according to claim 1, further comprising: a model update unit that updates the second storage unit and the correspondence table storage unit according to the constructed second estimation model.
A sensor abnormality data storage unit that accumulates sensor abnormality data in which the cause of abnormality of the facility is given to sensor data when the facility has an abnormality;
A model evaluation unit for obtaining an evaluation result of the first estimation model based on the sensor abnormality data;
An external request acquisition unit for acquiring external requests;
A model construction unit for constructing a first estimated model based on the evaluation result of the model evaluation unit and the data of the user request acquisition unit;
The abnormality cause estimation device according to claim 1, further comprising: a model update unit that updates the first storage unit according to the constructed first estimation model.
A data acquisition unit for acquiring sensor data of a sensor installed in the facility; a first estimation unit for obtaining a first estimation cause by a first estimation model for estimating the cause of abnormality of the facility; and the first A correspondence table storage unit that stores a correspondence table that correlates the estimated cause of the second estimated model supplementing the estimated cause, a second estimating unit that obtains a second estimated cause based on the second estimated model, An abnormality cause estimation method in an abnormality cause estimation device comprising:
The sensor data is acquired by the data acquisition unit,
In the first estimation unit, the first estimation model estimates the first estimation cause based on the sensor data,
In the second estimation unit, when the first estimation cause and the second estimation model are associated with each other based on the correspondence table stored in the correspondence table storage unit, the second estimation model Is an abnormal cause estimation method for estimating a second estimated cause.