JP6620402B2 - Event prediction system, event prediction method and program - Google Patents

Description

  The present invention relates to an event prediction system, an event prediction method, and a program.

  In a device that operates regularly, such as a gas engine, once a serious failure occurs, the device may be stopped for a long period of time. In this case, a large loss is caused to a user who uses the device. If abnormalities can be predicted in advance, equipment maintenance workers can perform preventive maintenance that can be operated without long-term outages by performing preliminary maintenance on planned shutdown days (and days with fewer operations). It becomes possible.

  For example, Patent Literature 1 learns a correlation between operation data of a device and a threshold value for determining an abnormality by data mining, and predicts an abnormality of the target device from the correlation and operation data of the target device. There is a disclosure about the system.

JP 2011-192097 A

However, in the method described in Patent Document 1, the predicted value is either abnormal or normal, and the prediction accuracy is unknown. Therefore, if the prediction probability is not very high, it is not suitable for operation, and as a result, it is likely to be suitable only for prediction immediately before the occurrence of an abnormality with high accuracy, and the cases applicable to preventive maintenance are limited. There is a problem.
Moreover, it is required not only to predict failures and abnormalities but also to predict the occurrence by paying attention to a specific event. The occurrence of an event can be defined as exceeding a threshold value or range set for a specific state quantity. For example, a state such as steady operation or high efficiency operation can be regarded as an event. Predicting these events is useful for studying appropriate operation methods and high-efficiency control methods for equipment and realizing energy savings and productivity improvements.

  Accordingly, an object of the present invention is to provide an event prediction system, an event prediction method, and a program that can solve the above-described problems.

According to a first aspect of the present invention, an operation data acquisition unit that acquires operation data indicating an operation state of a device to be predicted, operation data in a predetermined period, and an apparatus that acquired the operation data in the predetermined period A prediction model for predicting whether or not the event is likely to occur in the device in a predetermined prediction target period based on learning data associated with information indicating whether or not an event has occurred An operation state prediction unit that predicts whether or not the event occurs in a prediction target device based on the prediction model and the operation data acquired by the operation data acquisition unit, A prediction reliability calculation unit that calculates the reliability of the prediction based on the prediction of the operating state prediction unit and a record of whether or not the event has occurred in the device with respect to the prediction; E Bei an output unit results of prediction by the state predicting unit and outputs the reliability of the result of the prediction, and the prediction model generator is actually the prediction period on the days for which the event has occurred based Information indicating that the event has occurred is associated with the operation data from the previous day to the date when the event actually occurred, and the prediction target from the date to be predicted toward the future Generating a prediction model that predicts whether or not an abnormality can occur within a period, and the operation state prediction unit predicts a value based on the operation data of the day and the prediction model for each day of the operation data And the predicted reliability calculation unit, for each day of the operation data, if the event has actually occurred within the prediction target period from that day to the future, the operation data for that day Vs. The first performance information indicating that the event has occurred is associated, and if the event has not actually occurred, the second performance information indicating that the event has not occurred is associated, The total value of the number of the operation data for each prediction value associated with the operation state prediction unit is associated with the first record information or the second record information that is equal to the prediction result by the operation state prediction unit. Based on the total value of the number of the operation data associated with the first performance information and the total value of the number of the operation data associated with the second performance information among the operation data. An event prediction system for calculating the reliability of the prediction .

In the second aspect of the present invention, the prediction model generation unit generates a plurality of prediction models for predicting whether or not the event is likely to occur, and the operating state prediction unit is configured to generate the plurality of predictions. Prediction is performed based on each of the models, and the prediction reliability calculation unit calculates the reliability for each combination of prediction values based on each of the plurality of prediction models.

The prediction model generation unit according to the third aspect of the present invention generates a prediction model as to whether or not the event can occur for each of a plurality of types of events, and the operating state prediction unit includes the prediction model Based on each of the above, the occurrence of the corresponding event among the plurality of types of events is predicted.

According to a fourth aspect of the present invention, there is provided a step of obtaining operation data indicating an operation state of a device to be predicted, an operation data in a predetermined period, and an event to be predicted in the predetermined period in the device that acquired the operation data. Generating a prediction model for predicting whether or not the event occurs in the device, based on learning data associated with information indicating whether or not the prediction model and the acquired A step of predicting whether or not the event is likely to occur in a device to be predicted in a predetermined prediction target period based on the operation data; and the event in the device for the prediction and the prediction. Calculating a reliability of the prediction based on a record of whether or not it has occurred, outputting a result of the prediction and a reliability of the result of the prediction; It has a, in the step of generating the prediction model, to actually the operation data of the day when the event occurs until the day actually said event only from the day before the prediction target time to the reference is generated In association with information indicating that the event has occurred, a prediction model for predicting whether or not an abnormality can occur within the prediction target period from the date to be predicted to the future is generated. In the step of predicting, for each day of the operation data, the operation data of that day is associated with a prediction value based on the prediction model, and in the step of calculating the reliability of the prediction, the day of the operation data If the event has actually occurred within the prediction target period from that day to the future, first performance information indicating that the event has occurred with respect to the operation data of the day If the event does not actually occur, the second performance information indicating that the event has not occurred is associated, and the operation for each predicted value associated in the predicting step Of the operating data in which the total value of the number of data is associated with the first performance information or the second performance information that is equal to the prediction result in the prediction step, the first performance information is associated. In the event prediction method, the reliability of the prediction is calculated based on the total value of the number of operating data and the total value of the number of operating data associated with the second performance information .

According to a fifth aspect of the present invention, the computer of the event prediction system is configured to acquire operation data indicating an operation state of a device to be predicted, operation data in a predetermined period, and an apparatus that acquired the operation data to the predetermined period. A prediction model for predicting whether or not the event occurs in the device in a predetermined prediction target period based on learning data associated with information indicating whether or not an event that is a prediction target has occurred A means for generating, a means for predicting whether or not the event is likely to occur in a device to be predicted based on the prediction model and the acquired operation data, the prediction and the device for the prediction A means for calculating the reliability of the prediction based on a record of whether or not the event has occurred, and outputs a result of the prediction and a reliability with respect to the result of the prediction Stage, to function as the means for generating a prediction model, actually the operation from just before the day the event the prediction target period based on the date of occurrence until the day actually said event has occurred A prediction model that associates information indicating that the event has occurred with data and predicts whether or not an abnormality can occur within the prediction target period from the prediction target date toward the future For each day of the operational data, the means for associating the operational data for that day with a predicted value based on the prediction model, and calculating the reliability of the prediction. For each day, if the event has actually occurred within the prediction target period from that day to the future, first performance information indicating that the event has occurred with respect to the operation data of the day Corresponding If the event does not actually occur, the operation data for each prediction value associated with the second performance information indicating that the event has not occurred and associated with the predicting unit. Among the operating data in which the first result information or the second result information that is equal to the result of prediction by the means for predicting is associated with the first result information. The program calculates the reliability of the prediction based on a total value of the number of operating data and a total value of the number of operating data associated with the second performance information .

  ADVANTAGE OF THE INVENTION According to this invention, the event prediction system which can grasp | ascertain the reliability of prediction can be obtained.

It is a schematic block diagram which shows an example of a structure of the event prediction system in 1st embodiment of this invention. It is a 1st figure which shows an example of the data used for prediction of abnormality of the gas engine in 1st embodiment of this invention. It is a figure explaining the reliability of prediction in a first embodiment of the present invention. It is a 2nd figure which shows an example of the data used for prediction of abnormality of the gas engine in 1st embodiment of this invention. It is a processing flow of the event prediction system in 1st embodiment of this invention. It is a 3rd figure which shows an example of the data used for prediction of abnormality of the gas engine in 1st embodiment of this invention. It is an example of the figure which totaled the abnormal prediction of the gas engine in 2nd embodiment of this invention, and the performance with respect to the prediction. It is the figure which showed an example of the output of the predicted value in 2nd embodiment of this invention.

<First embodiment>
Hereinafter, an event prediction system according to an embodiment of the present invention will be described with reference to FIGS.
In the first embodiment and the second embodiment, as an example of the “event prediction system” of the present invention, the prediction target event is “failure” or “abnormal”, and the prediction target device is “gas engine”. However, the application target of the present invention is not limited to these.
FIG. 1 is a schematic block diagram showing an example of the configuration of the event prediction system in the first embodiment of the present invention. In the present embodiment, the event prediction system is configured by one server device, for example. The failure prediction apparatus 10 in FIG. 1 is an example of an event prediction system. The failure prediction apparatus 10 predicts whether or not a failure will occur in the gas engine, and assists the user in making an appropriate maintenance plan for the gas engine. As illustrated in FIG. 1, the failure prediction apparatus 10 includes an operation data acquisition unit 11, a prediction condition acquisition unit 12, a prediction model generation unit 13, an operation state prediction unit 14, a prediction reliability calculation unit 15, and an output. Unit 16 and storage unit 17.

The operation data acquisition unit 11 acquires operation data 20 of the gas engine. The operation data is information including a daily state quantity for each gas engine (for example, an average value of the engine speed, generated power, cylinder temperature, etc. per day). Moreover, the operation data acquisition unit 11 acquires an abnormality record of whether or not an abnormality has occurred on that day for each gas engine. Further, the operation data acquisition unit 11 acquires the latest operation data 20 of the gas engine to be predicted.
The prediction condition acquisition unit 12 acquires conditions necessary for prediction of a gas engine failure input by a user.
The prediction model generation unit 13 generates a prediction model for predicting whether or not an abnormality occurs in the gas engine based on the operation data.
The operating state prediction unit 14 becomes a prediction target during a predetermined prediction period based on the prediction model generated by the prediction model generation unit 13 and the latest operation data 20 of the gas engine acquired by the operation data acquisition unit 11. Predict whether an abnormality will occur in the gas engine.
The prediction reliability calculation unit 15 calculates the reliability of prediction based on the prediction model based on the prediction of the operating state prediction unit 14 and the actual results for the prediction.
The output unit 16 outputs the prediction reliability calculated by the operation state prediction unit 14 and the prediction reliability calculation unit 15.
The storage unit 17 stores various information such as operation data acquired by the operation data acquisition unit 11.

FIG. 2 is a first diagram illustrating an example of data used for predicting abnormality of the gas engine in the first embodiment of the present invention.
FIG. 2 is prediction history data including daily operation data of a certain gas engine A, an abnormal performance flag indicating whether or not an abnormality has occurred on that day, and a predicted value by the operating state prediction unit 14. In “Date”, a date indicating what time of operation data is stored. In the “cylinder temperature”, an average value of the temperature of one cylinder among a plurality of cylinders provided in the gas engine A is stored. For convenience of explanation, only the “cylinder temperature” is described as the state quantity of the gas engine, but the operation data includes items such as the average temperature of other cylinders, the engine speed, and the generated power. It may be. In the “predicted value”, a predicted value indicating whether or not an abnormality occurs on the day stored in the “date” item by the operating state predicting unit 14 is stored. For example, the predicted value is calculated based on the prediction model generated by the prediction model generation unit 13 based on the operation data up to the previous day and the abnormality record flag and the latest operation data of the day. It is a value that predicts whether or not an abnormality will occur within 24 hours at 0 o'clock. The “abnormal performance flag” stores a value indicating whether an abnormality has occurred in the gas engine A on the date stored in the “date” item. For example, prediction history data for two to three years is recorded in the storage unit 17.

The data in FIG. 2 will be described with reference to an example. In the data in the first row, the average cylinder temperature on January 1 is 500 ° C., and the operating state prediction unit 14 does not generate any abnormality on January 1 (predicted value). = "Normal"), and as a result, no abnormality has occurred (abnormal performance flag = "0").
Similarly, in the case of the data in the sixth row, the cylinder average temperature on January 6 is 600 ° C., and the operating state prediction unit 14 predicts that an abnormality will occur on January 6 (predicted value = “abnormal”). As a result, it indicates that an abnormality has occurred (abnormal performance flag = “1”).

  January 7 indicated by the data in the seventh row is, for example, the current date. The value of “cylinder temperature” in the seventh row is the latest cylinder temperature acquired by the operation data acquisition unit 11 on January 7. The value of “predicted value” in the seventh line (= “normal”) is determined based on the predicted model generated by the predicted model generating unit 13 and the latest cylinder temperature by the operating state predicting unit 14 on January 7th. It shows that it is predicted that no abnormality will occur in engine A.

The outline of the prediction process will be described. The “date”, “cylinder temperature” (operation data) and “abnormal performance flag” exemplified in FIG. 2 are learning data with the value of the “abnormal performance flag” as a label. The prediction model generation unit 13 generates a prediction model (predictor) for predicting the occurrence of abnormality by machine learning based on the learning data. As a machine learning method, for example, SVM (Support Vector Machine), decision tree, neural network, or the like can be used.
Next, the operation state prediction unit 14 acquires the latest operation data of the gas engine (in this case, the latest temperature of the cylinder on January 7), and the latest operation data and the prediction model generation unit 13 generate it. Whether or not an abnormality occurs in the gas engine is predicted based on the predictor. However, the prediction by the predictor generated by machine learning only predicts whether the state of the gas engine A is normal or abnormal, and does not know how reliable the prediction is. Therefore, in the present embodiment, the prediction reliability calculation unit 15 calculates the prediction reliability of the operating state prediction unit 14. Next, a method of calculating the prediction reliability by the prediction reliability calculation unit 15 will be described with reference to FIG.

FIG. 3 is a diagram for explaining the reliability of prediction in the first embodiment of the present invention.
FIG. 3 (a) is a table in which predictions for a given period and results for the predictions for a certain gas engine A are tabulated. Data in the first row of this table indicated by reference numerals 31 to 33 indicate the actual results and the reliability of prediction when the operation state prediction unit 14 predicts that an abnormality will occur. The value in the column 31 is the number of times that an abnormality has actually occurred when the operating state prediction unit 14 predicts that an abnormality will occur. The value in the column 32 is the number of times that the operating state prediction unit 14 predicted that an abnormality occurred, but no abnormality actually occurred. The value in the column 33 is the rate at which the operating state prediction unit 14 predicts that an abnormality will occur and the abnormality has actually occurred. Specifically, in the data on the first line, the operation state prediction unit 14 predicts that an abnormality occurs 100 times for the operation of the gas engine A in a predetermined period, of which 58 times an abnormality actually occurs and 42 times Indicates that there was no abnormality and it was operating normally. In addition, the abnormality occurrence rate at this time is 58%. In this case, it can be considered that the reliability with respect to the fact that the operation state prediction unit 14 has predicted that an abnormality will occur is 58%.

  On the other hand, the data in the second row indicated by reference numerals 34 to 36 indicate the results and the reliability of prediction when the operation state prediction unit 14 is predicted to operate normally. The value in the column 34 is the number of times that an abnormality has actually occurred, although the operating state prediction unit 14 is predicted to operate normally. The value in the column 35 is the number of times that the operation state prediction unit 14 is predicted to operate normally and actually operates normally. The value in the column 36 is a rate at which the operating state prediction unit 14 is predicted to operate normally, but an abnormality has actually occurred. Specifically, the data in the second row predicts that the operation state prediction unit 14 normally operates 1000 times for the operation of the gas engine A for a predetermined period, of which 100 times are actually abnormal and 900 Times indicate normal operation. In addition, the occurrence rate of abnormality at this time is 10%. In this case, the reliability with respect to the prediction that the operation state prediction unit 14 is normal can be considered to be 90%.

  FIG. 3B summarizes the operation data illustrated in FIG. 2 and represents the same as in FIG. In the case of the operation data in FIG. 2, from January 1 to January 4, the predicted value is normal and the actual operation result is 0 (normal). In addition, the predicted value from January 5th to January 6th is abnormal, and the actual result is 0 (normal) on January 5th, and 1 (abnormal) is predicted on January 6th. It is. The predicted reliability calculation unit 15 calculates the items shown in FIG.

  When the operation state prediction unit 14 predicts that the gas engine operation during the prediction period = January 7 (0 to 24:00) is “normal”, the output unit 16 calculates the prediction reliability together with the prediction value “normal”. An abnormality occurrence rate by the unit 15 (in this example, “0%”) is output. The user sees the output result by the output unit 16 (predicted value = normal, abnormality occurrence rate = 0%), and no abnormality occurs in the gas engine A on January 7, and the reliability of the prediction is 100. % Can be obtained. Also, if the operating state prediction unit 14 predicts “abnormal”, the user sees the output result by the output unit 16 (predicted value = abnormal, abnormality occurrence rate = 50%). It is possible to grasp that an abnormality occurs with a probability of 50% and to take measures and preparations for the occurrence of the abnormality.

So far, using FIG. 2 to FIG. 3 as an example, based on past learning data, an example of predicting whether or not an abnormality will occur with respect to operation on that day and calculating the reliability of the prediction will be described. Went.
However, it is preferable to perform preventive maintenance based on a long-term maintenance plan. Next, the prediction process of the failure prediction apparatus 10 according to the present embodiment will be described by taking as an example the case where the occurrence of an abnormality is predicted for a period of one month ahead.

FIG. 4 is a second diagram showing an example of data used for prediction of abnormality of the gas engine in the first embodiment of the present invention.
Compared to the data in FIG. 2, the prediction history data in FIG. 4 has a “predicted value” that is “predicted value for a period up to one month ahead” and an “abnormal performance flag (for prediction up to one month ahead). ) ”And“ abnormal performance flag (result up to one month ahead) ”. In the case of FIG. 4, it is assumed that the present day is February 4 and the operation data up to February 3 and the presence / absence of abnormality are known as results.
The “abnormal record flag (for prediction up to one month ahead)” stores an abnormal record flag used when the prediction model generation unit 13 predicts whether an abnormality will occur in the period up to one month ahead. . For example, “1” (abnormal) is stored in the “abnormal performance flag (for prediction up to one month ahead)” for the period one month before the date of actual occurrence of the abnormality (February 3). . For other periods, the same value as the abnormality record flag is stored. In the case of the example of FIG. 4, an abnormality actually occurred on February 3, and “1” is stored in the “abnormal performance flag” on February 3. It is assumed that no abnormality has occurred from January 5 to January 31 (not shown). In that case, the prediction model generation unit 13 sets 1 to the value of the “abnormal performance flag (for prediction up to one month ahead)” for the period up to one month before, based on February 3rd. That is, “1” is set in the “abnormal performance flag (for prediction up to one month ahead)” from January 3 to February 3. The prediction model generation unit 13 sets “0” (normal) for January 1 and January 2 as actual results. When performing prediction up to one month ahead, the prediction model generation unit 13 includes “date”, “cylinder temperature”, and “abnormal performance flag (for prediction up to one month ahead)” in the prediction history data of FIG. Machine learning is performed using the value to generate a prediction model (predictor). That is, the prediction model generation unit 13 generates a prediction model for predicting whether the state of the gas engine A is in a state at any point in the period from the time when the abnormality actually occurs until one month ago. Generate. In other words, the prediction model generation unit 13 abnormalizes the state of the gas engine A indicated by the operation data in a period from one month before the time when the abnormality actually occurs until the time when the abnormality actually occurs. A prediction model is generated assuming that a state in which the occurrence of the occurrence is possible. Moreover, when the operation state prediction unit 14 predicts that an abnormality can occur within one month based on the prediction model and the latest operation data, the operation state prediction unit 14 predicted that an abnormality will occur within one month ahead. It is considered. The “predicted value for the period up to one month ahead” in FIG. 4 describes the value predicted by the operating state predicting unit 14 in this way.

Next, the “abnormal performance flag (result up to one month ahead)” will be described. The value of “abnormal performance flag (result up to one month ahead)” is calculated when the prediction reliability calculation unit 15 calculates the reliability of prediction when the operating state prediction unit 14 performs prediction up to one month ahead. It is an item used for.
The value of “abnormal performance flag (actual up to one month ahead)” on January 1 indicates whether or not an abnormality has occurred between January 1 and one month. If an abnormality occurs, the value of this item is “1”, and if no abnormality has occurred, the value of this item is “0”. In the case of FIG. 4, it is assumed that no failure has occurred once in a period (January 5 to January 31) in which the description is omitted. Then, since no abnormality has occurred from January 1 to February 1, the prediction reliability calculation unit 15 sets “0” to the value of this item on January 1 as illustrated. The same applies to January 2nd. Next, considering January 3, since an abnormality has occurred on February 3, one month after January 3, the predictive reliability calculation unit 15 determines that “ “1” is set to the value of the “result flag (result up to one month ahead)”. Next, considering January 4, it has not been completed on February 4 (current) one month ahead. Therefore, the actual result with respect to the “predicted value for the period up to one month ahead” predicted on January 4 until today ends is unknown. However, in the case of FIG. 4, since an abnormality has already occurred on February 3, a result has been obtained as to whether an abnormality has occurred in the period from January 4 to one month ahead. Even if one month has not elapsed since the day when the prediction was made in this way, if an abnormality has occurred even once in the meantime, “1” is displayed as the “abnormal performance flag (result up to one month ahead)”. ”(Abnormal) can be made. The prediction reliability calculation unit 15 refers to the value of the “abnormal record flag” in the period from the date when the prediction is made to one month ahead, and the “abnormal record” corresponding to the “predicted value for the period up to one month ahead”. Set the value of “Flag (actual up to 1 month ahead)”.

  Next, the prediction reliability calculation unit 15 uses the data in which the “abnormal performance flag (result up to one month ahead)” can be set, and the “predicted value for the period up to one month ahead” of the data of each date. The data with the “abnormal performance flag (result up to one month ahead)” is classified, and the number of data is totaled. Specifically, the predictive reliability calculation unit 15 sets (1) “normal” as a set of “predicted value for a period up to one month ahead” and “abnormal performance flag (result up to one month ahead)”. ”And data that did not cause any abnormalities during the period up to one month ahead, (2) data that was predicted as“ normal ”and actually occurred during the period up to one month ahead, and (3)“ abnormal ” ”And data that did not cause any abnormalities during the period up to one month ahead, and (4) data that was predicted as“ abnormal ”and actually occurred during the period up to one month ahead. , (1) to (4), the number of data classified into each is totaled. Next, the prediction reliability calculation unit 15 predicts “abnormal” from these aggregation results, and the rate of occurrence of abnormality in the period up to one month ahead, and the abnormality is predicted in the period up to one month ahead predicted “normal”. And the rate at which occurrence occurred.

Next, the flow of failure prediction processing according to the present embodiment will be described with reference to FIG.
FIG. 5 is a processing flow of the event prediction system in the first embodiment of the present invention.
As an example, the case where it is predicted whether an abnormality will occur in the operation of the gas engine A in the period up to one month ahead will be described as an example. As a premise, the prediction history data illustrated in FIG. 4 is recorded in the storage unit 17. The prediction history data is associated with operation data related to the gas engine A for a sufficient period (for example, three years) and each operation data. It is assumed that a stored abnormal result flag and a predicted value predicted in the past (“predicted value for a period up to one month ahead”) are stored. For convenience of explanation, the state quantity of the gas engine used for prediction included in the operation data is only the average temperature of the cylinder.

  First, the operation data acquisition unit 11 acquires the latest operation data (step S11). For example, a temperature sensor is provided in the cylinder of the gas engine A, and the operation data acquisition unit 11 acquires the temperature of the cylinder at a predetermined time interval. The operation data acquisition unit 11 records the latest temperature in the storage unit 17. Next, the prediction condition acquisition unit 12 receives an input of prediction conditions and a prediction execution instruction from the user (step S12). The prediction condition is, for example, a prediction period (in this example, one month), identification information of a gas engine to be predicted (in this example, gas engine A), or a prediction model generated by any machine learning method. Or do it. The prediction condition acquisition unit 12 records the received prediction condition in the storage unit 17. The prediction condition acquisition unit 12 outputs the prediction condition to the prediction model generation unit 13 and instructs generation of a prediction model.

  The prediction model generation unit 13 instructed to generate the prediction model first performs preprocessing necessary for generation of the prediction model (step S13). The preprocessing is, for example, to set “abnormal performance flag (for prediction up to one month ahead)” in the prediction history data illustrated in FIG. The prediction model generation unit 13 sets the value of the “abnormal performance flag” on the same day to the “abnormal performance flag (for prediction up to one month ahead)” of the prediction history data recorded in the storage unit 17. Next, the prediction model generation unit 13 sets “1” indicating an abnormality to the value of the “abnormal performance flag (for prediction up to one month ahead)” for the period one month before the day when the abnormality actually occurred. . The above is the preprocessing.

  Next, the prediction model generation unit 13 generates a prediction model (predictor) (step S14). For example, when the user instructs to generate a predictor by SVM in step S12, the prediction model generation unit 13 sets the operation data and the value of “abnormal performance flag (for prediction up to one month ahead)”. Based on the combined learning data, the function of the boundary between “abnormal performance flag (for prediction up to one month ahead)” = 0 (normal) and “abnormal performance flag (for prediction up to one month ahead)” = 1 (abnormal) calculate. The calculated boundary function is a predictor. Next, the prediction model generation unit 13 outputs information of the calculated function to the operating state prediction unit 14 and instructs to predict whether or not an abnormality will occur in the gas engine A in the period up to one month ahead. The operating state prediction unit 14 predicts whether or not an abnormality occurs in the gas engine A based on the acquired boundary function (step S15). Specifically, the operation state prediction unit 14 determines whether the latest temperature of the cylinder acquired by the operation data acquisition unit 11 is normal or abnormal based on a boundary function. When it is determined that the latest temperature is normal, the operation state prediction unit 14 predicts that the operation state of the gas engine up to one month ahead is normal. When it is determined that the latest temperature is abnormal, the operating state prediction unit 14 predicts that an abnormality will occur in the gas engine in the period up to one month ahead. The operating state prediction unit 14 outputs the predicted value to the output unit 16 and instructs the prediction reliability calculation unit 15 to calculate the abnormality occurrence rate.

  Next, the predicted reliability calculation unit 15 calculates an abnormality occurrence rate (step S16). First, the prediction reliability calculation unit 15 refers to the value of the “abnormal performance flag” within one month from that day for each day, and if there is any data having “1” as a value, “1” is set to the value of “abnormal performance flag (result up to one month ahead)”, and “0” is set to “abnormal performance flag (result up to one month ahead)” on other days. Next, the prediction reliability calculation unit 15 combines data “predicted value for a period up to one month ahead” and “abnormal performance flag (actual up to one month ahead)” recorded in advance in association with learning data. Are classified into the following four, and the number of data classified into each of them is totaled. (1) “Predicted value for period up to one month ahead” = normal and “abnormal performance flag (result up to one month ahead)” = 0, (2) “Predicted value for period up to one month ahead” = Normal and “abnormal performance flag (result up to one month ahead)” = 1, (3) “predicted value for period up to one month ahead” = abnormal and “abnormal performance flag (result up to one month ahead)” = 0 data, (4) “predicted value for period up to 1 month ahead” = abnormal and “abnormal performance flag (actual up to 1 month ahead)” = 1. Next, the prediction reliability calculation unit 15 calculates an abnormality occurrence rate from the collected values. Specifically, the prediction reliability calculation unit 15 calculates the “predicted value for the period up to one month ahead” = the abnormality occurrence rate for normality by (total value of (2) ÷ [total value of (1) + (2) [Aggregated value of]], and the “predicted value for the period up to one month ahead” = abnormality rate for abnormalities is calculated by (4) aggregated value [(3) aggregated value + (4) aggregated value] calculate. The prediction reliability calculation unit 15 outputs the calculated abnormality occurrence rate to the output unit 16.

  The output unit 16 selects an abnormality occurrence rate corresponding to the predicted value acquired from the operating state prediction unit 14. For example, when the operation state prediction unit 14 predicts that “the operation state of the gas engine up to one month ahead is normal”, the output unit 16 sets “the predicted value for the period up to one month ahead” = the abnormality occurrence rate for normality. select. Similarly, when the operation state prediction unit 14 predicts that “an abnormality occurs in the gas engine in a period up to one month ahead”, the output unit 16 outputs “an estimated value for a period up to one month ahead” = an abnormality occurrence for the abnormality. Select a rate. Next, the output unit 16 outputs and displays the predicted value acquired from the operating state prediction unit 14 and the selected abnormality occurrence rate on a display device connected to the failure prediction device 10 (step S17).

  According to the present embodiment, the user predicts whether the abnormality occurs in the gas engine within a predetermined prediction period and whether the prediction based on past operation data is a highly reliable prediction or a low reliability prediction. It is possible to grasp the above, and it is possible to make a maintenance plan in consideration of the probability of occurrence of abnormality.

Next, an example in which the user makes a maintenance plan in consideration of the severity of the failure and the reliability of the prediction will be described with reference to FIG.
FIG. 6 is a third diagram showing an example of data used for prediction of abnormality of the gas engine in the first embodiment of the present invention.
FIG. 6 shows prediction history data including predicted values and actual results for light and heavy failures occurring within one month. A minor failure means that, for example, one of a plurality of cylinders provided in a gas engine fails. A serious failure means that three or more of a plurality of cylinders included in a gas engine fail. Stored in the “predicted value for minor failure in a period up to one month ahead” is a predicted value for failure of one of the cylinders within one month. The “predicted value for a serious failure in a period up to one month ahead” stores a predicted value for the simultaneous failure of three or more cylinders within one month. “Minor failure record flag” and “Major failure record flag” respectively describe whether a minor failure has occurred on that day and whether a major failure has occurred on that day.

  “Minor failure record flag (for prediction up to one month ahead)” and “Severe failure record flag (for prediction up to one month ahead)” are labels of learning data for the prediction model generation unit 13 to generate a predictor. It is. The values of “Minor failure record flag (for prediction up to one month ahead)” and “Major failure record flag (for prediction up to one month ahead)” are values set in the pre-processing of step S13 in FIG. is there. For example, according to the “Minor Failure Result Flag” on February 2 in FIG. 6, a minor failure has occurred on February 2. Accordingly, the prediction model generation unit 13 sets “1” to the “Minor failure record flag (for prediction up to one month ahead)” from January 2 to February 2.

  “Minor failure record flag (results up to one month ahead)” and “Major failure record flag (results up to one month ahead)” indicate whether a minor failure actually occurred in the period up to one month ahead or Each value indicating whether a failure has occurred is stored. This value is the value set by the prediction reliability calculation unit 15 in step S16 of FIG. For example, according to FIG. 6, a minor failure has occurred on February 2, and a minor failure has occurred within one month from each day in the period from January 2 to February 2. Therefore, the predicted reliability calculation unit 15 sets 1 to the value of “Minor failure record flag (results up to one month ahead)” from January 2 to February 2.

  As described above, the prediction model generation unit 13 is based on the “date”, “cylinder temperature”, and “light failure record flag (for prediction up to one month ahead)” of the prediction history data in FIG. A predictor for predicting a minor failure is generated by machine learning. Further, the prediction model generation unit 13 detects a serious failure by machine learning based on the “date”, “cylinder temperature”, and “major failure record flag (for prediction up to one month ahead)” of the prediction history data in FIG. Generate a predictor for prediction. The operating state prediction unit 14 predicts whether or not a light failure will occur within one month by a predictor for predicting a light failure, and the predicted value is displayed in “period up to one month ahead” in FIG. Set to “Predicted value for minor failure”. The operating state prediction unit 14 predicts whether or not a serious failure will occur within one month by a predictor for predicting a serious failure, and the predicted value is displayed in the period up to one month ahead in FIG. Set to “Predicted value for serious failure”.

  The predictive reliability calculation unit 15 classifies data in which the “predicted value for a light failure in a period up to one month ahead” and the “light failure record flag (result of a light failure up to one month ahead)” in FIG. And calculate the failure rate for light failure prediction. Further, the predictive reliability calculation unit 15 sets a combination of the “predicted value for a serious failure in the period up to one month ahead” and the “major failure result flag (the result of serious failure up to one month ahead)” in FIG. Are classified and tabulated, and the failure occurrence rate for the prediction of serious failure is calculated.

  The output unit 16 outputs a prediction and failure occurrence rate for a minor failure, and a prediction and failure occurrence rate for a major failure. For example, it is assumed that the output result by the output unit 16 is “minor failure: abnormal, failure occurrence rate: 20%” and “serious failure: normal, failure occurrence rate: 10%”. From this output result, the user may experience a minor failure within one month, but the prediction reliability is 20%, and no serious failure occurs within one month, and the prediction reliability is 90%. %, And for example, it is decided to postpone the maintenance of the gas engine A. For example, it is assumed that the output result by the output unit 16 is “minor failure: abnormal, failure occurrence rate: 70%” and “serious failure: normal, failure occurrence rate: 10%”. For example, the user compares the cost when a minor failure occurs with the maintenance cost, and if the maintenance cost seems to be higher, the user does not plan maintenance until the failure occurrence rate reaches 90% for the minor failure. Make a decision. For example, it is assumed that the output result by the output unit 16 is “minor failure: normal, failure occurrence rate: 10%” and “serious failure: abnormal, failure occurrence rate: 40%”. For example, when it is predicted that a serious failure will occur, the user performs maintenance of the gas engine A if, for example, the failure occurrence rate is 20% or higher, even if the reliability of the prediction is not so high. To plan.

  According to this example, it is possible to devise a maintenance plan that is more suitable for the situation according to the severity of the failure and the reliability of the prediction.

<Second embodiment>
Hereinafter, an event prediction system according to a second embodiment of the present invention will be described with reference to FIGS.
In the first embodiment, prediction is performed using one predictor, and the reliability for the prediction is calculated. In the second embodiment, prediction is performed using a plurality of predictors, and the reliability of prediction is calculated for each combination of predicted values by each predictor.

FIG. 7 is an example of a diagram in which the abnormality prediction of the gas engine in the second embodiment of the present invention and the results for the prediction are tabulated.
In the table of FIG. 7, when three types of predictors are provided, the results of the prediction and the abnormality occurrence rate are described for all combinations of predictions performed by the respective predictors. For example, in the data on the first row, the gas engine to be predicted has 100 times predicted to be abnormal by any of the predictor 1, the predictor 2, and the predictor 3, and an abnormality actually occurred. The number of times is 90, and the number of times that no abnormality occurred is 10 times. In this case, the abnormality occurrence rate is 90%. The same applies to the data after the second line.

  The processing of this embodiment will be described. For example, a machine learning method used for generation of the predictors 1 to 3 is recorded in the storage unit 17 in advance, and the prediction model generation unit 13 reads out the recording and performs each method. Predictors 1 to 3 are generated based on the learning data (operation data and abnormality record flag). Next, the operation state prediction unit 14 operates a gas engine in a predetermined prediction period from the plurality of predictors 1 to 3 generated by the prediction model generation unit 13 and the latest operation data acquired by the operation data acquisition unit 11. Whether or not a failure occurs is predicted, and the predicted value is recorded in the storage unit 17. Next, the prediction reliability calculation unit 15 associates the predicted values by the plurality of predictors 1 to 3 with the results for the prediction in the same manner as in the first embodiment. Moreover, the prediction reliability calculation part 15 totals for every result of whether abnormality generate | occur | produced with respect to all the combinations of the predicted value of each predictor. For example, the prediction reliability calculation unit 15 predicts the operation period of the gas engine when the predicted value by the predictor 1 is abnormal, the predicted value by the predictor 2 is abnormal, and the predicted value by the predictor 3 is abnormal. The number of times that no abnormality has occurred once (for example, one month) is counted. The same calculation is performed for other combinations of predicted values. When the aggregation is performed, the prediction reliability calculation unit 15 calculates the abnormality occurrence rate for each combination of the prediction values and records it in the storage unit 17. The output unit 16 outputs a prediction value by each predictor and an abnormality occurrence rate for the combination of the prediction values.

FIG. 8 is a diagram showing an example of output of predicted values in the second embodiment of the present invention.
In this output result, a predicted value based on the latest operation data by each predictor is displayed. In addition, the abnormality occurrence rate for the displayed combination of predicted values is displayed.
The user can know that an abnormality will occur within one month by the predictor 1 and the predictor 3 by looking at the output result, and the reliability of the prediction is 60%.
According to the present embodiment, it is possible to obtain an abnormality occurrence rate for each combination of predictions by a plurality of predictors, and the user can determine an output result based on more detailed prediction reliability.

  Here, for example, when the predictor 1 can only predict a short-term destination but has high accuracy, the predictor 2 can predict a long-term destination but has low accuracy, and the predictor 3 uses other predictors. By combining these, it is possible to effectively use when predictions for a plurality of prediction periods are performed. For example, when predicting a short-term destination, the predicted value of the predictor 2 may become noise. In such a case, the prediction reliability calculation unit 15 adds the prediction value when the prediction by the predictor 2 is excluded in addition to the result when the prediction values by all the predictors as illustrated in FIG. 7 are combined. A list may be generated, and the abnormality occurrence rate in each case may be output.

  Further, for example, the predictor 1 may be generated as a prediction model by SVM, the predictor 2 may be generated by a neural network, and the predictor 3 may be generated by a decision tree. SVMs and neural networks are known to be able to make accurate predictions, but they do not indicate why such predictions were made. On the other hand, prediction using a decision tree has an advantage that it presents the rule that led to the prediction, although the prediction accuracy is reduced. The rule that led to the prediction is, for example, a matter such as “the frequency at which the average temperature of the cylinder is X ° C. and the engine speed is Y rpm is Z times / week” as a basis for predicting that an abnormality will occur. . The user can refer to the maintenance work to be performed with reference to this rule.

  According to the present embodiment, it is possible to obtain an abnormality occurrence rate for each combination of predictions by a plurality of predictors, and the user can determine an output result based on more detailed prediction reliability.

  Each process in the event prediction system described above is stored in a computer-readable recording medium in the form of a program, and the program is read out and executed by the computer of the event prediction system. Is called. Here, the computer-readable recording medium means a magnetic disk, a magneto-optical disk, a CD-ROM, a DVD-ROM, a semiconductor memory, or the like. Alternatively, the computer program may be distributed to the computer via a communication line, and the computer that has received the distribution may execute the program.

The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.
In addition, the event prediction system may be configured by a single computer or may be configured by a plurality of computers connected so as to be communicable.

  In addition, it is possible to appropriately replace the components in the above-described embodiments with known components without departing from the spirit of the present invention. The technical scope of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the spirit of the present invention. For example, the event to be predicted is not “failure” or “abnormal”, but any event such as “whether the device to be predicted enters a high-efficiency operation state” or “a certain state quantity reaches a predetermined threshold” It may be a prediction for the general public. The failure prediction apparatus 10 is an example of an event prediction system. A gas engine is an example of equipment. A minor failure and a major failure are examples of multiple types of events.

DESCRIPTION OF SYMBOLS 10 ... Failure prediction apparatus 11 ... Operation data acquisition part 12 ... Prediction condition acquisition part 13 ... Prediction model generation part 14 ... Operation state prediction part 15 ... Prediction reliability calculation part 16. ..Output unit 17 ... storage unit

Claims (5)

An operation data acquisition unit that acquires operation data indicating the operation state of the device to be predicted;
Based on learning data in which operation data in a predetermined period and information indicating whether an event to be predicted in the predetermined period has occurred in the device that acquired the operation data are associated with the predetermined prediction target period A prediction model generation unit for generating a prediction model for predicting whether or not the event may occur in the device;
Based on the prediction model and the operation data acquired by the operation data acquisition unit, an operation state prediction unit that predicts whether or not the event occurs in a device to be predicted;
A prediction reliability calculation unit that calculates the reliability of the prediction based on the prediction of the operating state prediction unit and a record of whether or not the event has occurred in the device for the prediction;
An output unit that outputs a result of prediction by the operating state prediction unit and a reliability for the result of the prediction;
Bei to give a,
The prediction model generation unit generates the event with respect to the operation data from the day before the prediction target period to the date when the event actually occurred based on the day when the event actually occurred In association with the information indicating that it has been, generate a prediction model that predicts whether or not an abnormality can occur within the prediction target period toward the future from the date to be predicted,
The operating state prediction unit associates the operating data of the day with a prediction value based on the prediction model for each day of the operating data,
The predicted reliability calculation unit, for each day of the operation data, if the event has actually occurred within the prediction target period from that day to the future, The first performance information indicating that the event has occurred is associated, and if the event has not actually occurred, the second performance information indicating that the event has not occurred is associated with the operation. The total value of the number of the operation data for each prediction value associated with the state prediction unit and the first record information or the second record information that is equal to the prediction result by the operation state prediction unit are associated with each other Based on the total value of the number of the operational data associated with the first performance information and the total value of the number of the operational data associated with the second performance information among the operational data, Calculate prediction confidence
Event prediction system. The prediction model generation unit generates a plurality of prediction models in predicting whether or not the event can occur.
The operating state prediction unit performs prediction based on each of the plurality of prediction models,
The prediction reliability calculation unit calculates the reliability for each combination of prediction values based on each of the plurality of prediction models;
The event prediction system according to claim 1 . The prediction model generation unit generates a prediction model as to whether or not the event can occur for each of a plurality of types of events,
The operating state prediction unit predicts the occurrence of a corresponding event among the plurality of types of events based on each of the prediction models.
The event prediction system according to any one of claims 1 to 2 . Obtaining operational data indicating the operational status of the device to be predicted;
Based on learning data in which operation data in a predetermined period and information indicating whether an event to be predicted in the predetermined period has occurred in the device that acquired the operation data are associated with the predetermined prediction target period Generating a prediction model for predicting whether or not the device is ready for the event;
Predicting whether the event occurs in a prediction target device based on the prediction model and the acquired operation data;
Calculating the reliability of the prediction based on the prediction and a record of whether the event has occurred in the device for the prediction;
Outputting the result of the prediction and the reliability for the result of the prediction;
Have a,
In the step of generating the prediction model, the event is performed on the operation data from the day before the prediction target period to the date when the event actually occurred, based on the day when the event actually occurred. In association with information indicating that occurred, a prediction model for predicting whether or not an abnormality can occur within the prediction target period toward the future from the date to be predicted is generated,
In the step of predicting, for each day of the operation data, the operation data for that day is associated with a prediction value based on the prediction model,
In the step of calculating the reliability of the prediction, if the event has actually occurred within the prediction target period from that day to the future for each day of the operation data, the operation data for that day is included in the operation data. The first record information indicating that the event has occurred is associated with the second record information indicating that the event has not occurred if the event has not actually occurred. The total value of the number of the operation data for each predicted value associated in the predicting step is associated with the first record information or the second record information that is equal to the prediction result in the predicting step. Based on the total value of the number of the operation data associated with the first performance information and the total value of the number of the operation data associated with the second performance information among the operation data. The prediction Calculates the reliability,
Event prediction method. Event prediction system computer
Means for obtaining operating data indicating the operating status of the device to be predicted;
Based on learning data in which operation data in a predetermined period and information indicating whether an event to be predicted in the predetermined period has occurred in the device that acquired the operation data are associated with the predetermined prediction target period Means for generating a prediction model for predicting whether or not the device is ready for the event,
Means for predicting whether or not the event occurs in a prediction target device based on the prediction model and the acquired operation data;
Means for calculating a reliability of the prediction based on the prediction and a record of whether the event has occurred in the device for the prediction;
Means for outputting the result of the prediction and the reliability for the result of the prediction;
To function as,
The means for generating the prediction model is configured to generate the event with respect to the operation data from the date before the prediction target period to the date when the event actually occurred, based on the date when the event actually occurred. In association with information indicating that occurred, a prediction model for predicting whether or not an abnormality can occur within the prediction target period toward the future from the date to be predicted is generated,
The means for predicting associates the operation data of the day with a prediction value based on the prediction model for each day of the operation data,
The means for calculating the reliability of the prediction includes, for each day of the operation data, if the event has actually occurred within the prediction target period from that day to the future, The first record information indicating that the event has occurred is associated with the second record information indicating that the event has not occurred if the event has not actually occurred. The total value of the number of the operation data for each prediction value associated with the predicting means is associated with the first record information or the second record information that is equal to the prediction result by the predicting unit. Based on the total value of the number of the operational data associated with the first performance information and the total value of the number of the operational data associated with the second performance information among the operational data, Calculate the reliability of the prediction To,
program.

Images