WO2024042733A1

WO2024042733A1 - Failure probability evaluation system

Info

Publication number: WO2024042733A1
Application number: PCT/JP2023/001915
Authority: WO
Inventors: 将央渡辺; 寛新谷; 洋輔植木; 薫網本
Original assignee: 株式会社日立製作所
Priority date: 2022-08-24
Filing date: 2023-01-23
Publication date: 2024-02-29
Also published as: JP2024030392A

Abstract

The purpose of the present invention is to select failure-related time-series operation data from time-series operation data measured by a large quantity of sensors, and more accurately predict the remaining life of components constituting a mechanical system. The present invention is a failure probability evaluation system for evaluating the failure probability of components constituting a mechanical system, said failure probability evaluation system comprising: a maintenance history database that stores maintenance history data of the mechanical system; an operation database that stores a plurality of sets of operation data indicating the operation status of the components; a degree-of-dispersion calculation unit that, on the basis of the plurality of sets of operation data and the maintenance history data, calculates a degree of dispersion, which indicates the degree of variation of a failure probability function for calculating a failure probability for each set of operation data, and which corresponds to a correlation with failures of the components; and an operation data selection unit that selects operation data according to the calculated degree of dispersion. The failure probability evaluation system achieves the evaluation of the failure probability using, preferentially, the selected operation data.

Description

Failure probability evaluation system

The present invention relates to a technique for evaluating the failure probability of an object. Among these, it particularly relates to failure diagnosis and prediction (premonition), including calculation of failure probability. Here, the objects include equipment, facilities, mechanical systems, and parts that constitute these.

In various types of industrial machinery, such as power generation equipment and transportation equipment, which are a type of mechanical system, it is desired that the mechanical systems perform their specified functions normally. To this end, it is important to appropriately understand and manage the risk of failure of each part that makes up industrial machinery, and to perform maintenance such as repair or replacement of each part at an appropriate time. In addition, when managing and operating multiple industrial machines of the same type, statistical analysis of maintenance records that have occurred in the past can be used to determine the number of corrective maintenance and preliminary maintenance that may occur in the future, and the remaining life of the target equipment. It is possible to predict. Here, the maintenance record refers to data in which the content of maintenance and the time of occurrence are recorded as a pair.

In statistical analysis using maintenance records, failure probability density function f(t) and failure probability function F are required to predict the number of corrective maintenance and preliminary maintenance that may occur in the future, and the remaining life of the target equipment. (t), failure rate function λ(t), etc. Techniques for this purpose are disclosed in Non-Patent Document 1 and the like. Here, t is the operating time of the equipment. Note that when the failure probability density function f, the failure probability function F, and the failure rate function λ are functions of the variable t, the respective relationships can be expressed by (Equation 1) and (Equation 2).

Although Non-Patent Document 1 assumes that the operating status of a mechanical system is a static status, that is, there is no variation in space and time, the operating status of a mechanical system is generally not constant in space and time. For example, the operating status of a wind power generator changes from moment to moment depending on the wind conditions, and the load varies depending on the location conditions. In addition, the load on construction machinery and the like changes depending on the work environment, such as the work items that change from day to day, the driving characteristics of the driver, and the characteristics of the soil. Therefore, there is a limit to the accuracy of estimating the number of maintenance cycles and remaining life when evaluating a failure probability function or a failure rate function simply based on operating time.

In recent years, various sensors have been attached to mechanical systems, and it has become easy to access time-series operational data measured by these sensors via networks. Therefore, in order to improve the above-mentioned estimation accuracy, Patent Document 1 evaluates the operating status that differs for each individual from time-series operating data measured by sensors, and takes this into account appropriately. This makes it possible to estimate the number of maintenance cycles and remaining life.

JP2019-160128A

Here, in recent mechanical systems, it is not uncommon for a large number of sensors, ranging from hundreds to thousands, to be attached to the equipment. However, not all time-series operation data is necessarily related to failures, and a large amount of time-series operation data unrelated to failures is also included. Therefore, in order to further improve the accuracy of estimating the number of maintenance cycles and remaining service life, it is desirable to select in advance only time-series operation data related to failure and perform failure probability evaluation. However, Patent Document 1 does not disclose a technique for pre-selecting only time-series operation data related to a failure. Therefore, an object of the present invention is to further improve the accuracy of estimation regarding maintenance of an object, such as the number of maintenance cycles and remaining life.

In order to solve the above problems, the present invention calculates the failure probability using operation data (for example, time-series operation data). The present invention also includes a configuration that indicates the degree of dispersion of the failure probability function for calculating the failure probability, specifies the degree of dispersion according to the correlation with failure, and selects operation data accordingly. It will be done. More preferably, it is desirable to select operation data with a small degree of dispersion. A more specific configuration of the present invention includes, in a failure probability evaluation system for evaluating the failure probability of failure in parts constituting a mechanical system, a maintenance history database that stores maintenance history data of the mechanical system; An operation database that stores a plurality of operation data indicating the operation status of parts, and a degree of variation in a failure probability function for calculating a failure probability for each operation data based on the operation data and maintenance history data, a dispersion degree calculation section that calculates a dispersion degree according to the correlation with failures in the parts; and an operation data selection section that selects operation data according to the calculated dispersion degree; This is a failure probability evaluation system that implements evaluation of the failure probability by preferentially using data.

The present invention also includes a failure probability evaluation method using the failure probability evaluation system, a failure probability evaluation program that causes the failure probability evaluation system to function as a computer, and a storage medium that stores the same.

According to the present invention, it is possible to further improve the accuracy of estimation regarding maintenance of a target object compared to the conventional method. Problems, configurations, and effects other than those described above will be made clear by the description of the following examples.

System configuration diagram in Example 1 Diagram showing an example of maintenance history data used in Example 1 A diagram showing an example of time-series operation data and a division flag table stored in the time-series operation database 2 used in the first embodiment. Diagram showing an example of survival analysis data used in Example 1 A diagram showing the failure probability function and its coefficient of variation for the cumulative value of time-series operation data in Example 1 Flowchart for calculating the number of time-series operating data N _s in Example 1 in Example 1 Graph showing the relationship between the deviation rate R and the number of selected time-series operation data _Ns in Example 1 Graph showing the relationship between failure probability and cumulative damage in Example 1 System configuration diagram including a modified example of the failure probability evaluation unit in Example 1 A diagram showing an example of a graphical user interface (GUI) on the display unit in Example 1 System configuration diagram in Example 2 System configuration diagram in Example 3 Hardware configuration diagram of failure probability evaluation system in Example 4

Hereinafter, a failure probability evaluation system 100 according to an embodiment of the present invention will be described. A failure probability evaluation system 100 of the present embodiment includes a maintenance history database that stores maintenance history data of the mechanical system; An operation database that stores a plurality of operation data indicating the operation status of the component, and a degree of variation in a failure probability function for calculating a failure probability for each operation data based on the operation data and maintenance history data. , a dispersion degree calculation section that calculates a dispersion degree according to the correlation with failures in the parts, and an operation data selection section that selects operation data according to the calculated dispersion degree, The evaluation of the failure probability is realized by preferentially using operation data.

According to this embodiment, the failure probability of components of a mechanical system can be predicted with high accuracy by selecting operational data of sensors related to failure from a large amount of operational data and using it for failure probability evaluation. I can do it.

Hereinafter, each example showing more specific examples of this embodiment will be described using the drawings. The failure probability evaluation system 100 is realized on a computer, and its functions are executed by a processing device according to a program. However, the function may be realized by dedicated hardware, and the present invention is not limited to the use of a program (software).

Furthermore, in each of the following examples, the wind power generator 1 is used as a mechanical system, and the failure probability of the wind power generator 1 and its component parts is evaluated. However, the application of the present invention is not limited only to the wind power generator 1.

Further, in this embodiment, the above-mentioned (Equation 1) and (Equation 2) are used, but as is clear from these, it is necessary to identify any one of the failure probability density function f, the failure probability function F, and the failure rate function λ. If we can do this, we can calculate other functions. Therefore, in the following embodiments, when identifying the probability distribution of failures, the failure probability density function f and the failure probability function F are used depending on necessity.

FIG. 1 is a system configuration diagram in the first embodiment. The failure probability evaluation system 100 of this embodiment includes a maintenance history database 3, a time-series operation database 2, a time-series operation data selection section 13, and a failure probability evaluation section 10. Note that the time-series operation database 2 and the maintenance history database 3 may be provided outside the failure probability evaluation system 100, or may be configured as a single database.

First, the maintenance history database 3 and the maintenance history data 6 stored therein will be explained. In the maintenance history database 3 shown in FIG. 1, maintenance history data 6 of the components of the wind power generator 1 is accumulated. FIG. 2 is a diagram showing an example of maintenance history data 6 used in this embodiment. The maintenance history data 6, as shown in FIG. 2, is data in which individuals in which maintenance has occurred are associated with dates and times when maintenance has occurred. In this embodiment, the site name and machine number to which the wind power generator 1, which is the object, belongs are used as information for identifying an individual. Note that an individual number or the like may be used as information for identifying an individual. This also applies to each data described below. Furthermore, it is desirable that the maintenance history data 6 include a maintenance history including events (maintenance events) related to maintenance performed on the individual. As shown in FIG. 2, in this embodiment, each item of maintenance content, part name, and content is used as the maintenance history.

Here, the maintenance content indicates the type of maintenance performed, and in this embodiment, "pre-maintenance" and "post-maintenance" are used. "Preliminary maintenance" refers to maintenance such as replacing parts before a failure occurs, such as through periodic replacement. "Corrective maintenance" refers to maintenance such as replacing parts after an abnormality or failure occurs. Further, the component name identifies the component or location that has been maintained or where an abnormality has occurred. Additionally, the content indicates the content of the maintenance performed.

Additionally, the maintenance history data 6 may include the following items. For example, in the case of preliminary maintenance, the reason for carrying out the maintenance can be used. Furthermore, in the case of corrective maintenance, occurrence events of failures, etc. can be used.

In addition, if the wind power generator 1 is equipped with a function to automatically detect the maintenance history, the wind power generator 1 and the maintenance history database 3 are connected via a network, and the maintenance history data 6 is automatically transferred to the maintenance history database 3. May be accumulated. Alternatively, the worker 4 in charge of maintenance may register the maintenance history in the maintenance history database 3. With this configuration, maintenance history including maintenance details for a plurality of component parts is accumulated in the maintenance history database 3.

Next, the time-series operation database 2 and the time-series operation data 5 stored therein will be explained. In this embodiment, time-series operation data 5 such as operation data regarding the components of the wind power generator 1 is accumulated in a time-series operation database 2 via a communication means such as a network. At this time, although the collection interval of each data is not particularly limited, it can be set to the maximum interval within a range that does not lose the amount of information about the target physical phenomenon. In this case, it is possible to collect meaningful data for failure probability evaluation while reducing data volume. For example, in the failure probability evaluation of the wind power generator 1 according to the present embodiment, since relatively long-term predictions such as several months or several years are handled, an interval of approximately one day is ideal. Note that the time-series operation data 5 is a type of operation data.

Although the time-series operation data 5 may be measured values sampled at arbitrary intervals, it is more preferable to use statistical values such as the maximum value, minimum value, average value, and standard deviation within the collection interval. This makes it possible to make the most of information while significantly reducing the amount of data. For example, the average wind speed per day can be considered. Further, the information accumulated in the time-series operation database 2 is not limited to information obtained from the component parts themselves. For example, weather data such as temperature measured by weather observation equipment installed near the component is also useful in evaluating the load state of the component.

Here, FIG. 3 is a diagram showing an example of the time-series operation data 5 and the division flag table 41 stored in the time-series operation database 2 used in this embodiment. First, FIG. 3(a) is an example of the time-series performance data 5 output to the time-series performance data selection unit 13. As shown in FIG. 3(a), the time-series operation data 5 includes the following items: date and time, site name, machine number, and sensors 1 to N. The date and time indicates the date and time when measurements were taken by sensors 1 to N. The site name and machine number are the same as those in maintenance history data 6, and indicate the measurement targets of sensors 1 to N. Sensors 1 to N indicate sensors that measure time-series operation data 5, respectively. Here, in the wind power generator 1, the average wind speed is measured as the time-series operation data 5 of the sensor 1. In addition, other sensors 2 to N measure time-series operation data other than the average wind speed. That is, in this embodiment, N pieces of time-series operating data are measured by a total of N sensors 1 to N. In the example of FIG. 3(a), the data collection interval (measurement interval) is one day. Note that here, for convenience, information other than information obtained directly from component parts, such as temperature measured by weather observation equipment, is also referred to as a sensor.

Furthermore, prior to a detailed explanation of the division flag table 41 shown in FIG. 3(b), the necessity thereof will be described. Similar to general machine learning, failure probability evaluation devices need to have high prediction accuracy not only for the data used for learning but also for unknown data. In addition, in order to evaluate the prediction performance for unknown data, we divide the learning data used in the learning of the failure probability evaluation unit 10, which will be described later, and the test data used as a substitute for the unknown data, and calculate the failure probability for each. It is necessary to compare. Furthermore, in the division flag table 41, division flags are associated with each object (wind power generator 1). Specifically, in FIG. 3(b), the site name and machine number, which are examples of information for identifying the individual object, are used.

Therefore, in this embodiment, based on the division flag table 41, the time-series operation data 5 is assigned a division flag (learning or testing) according to the site name and machine number. As a result, the time-series operating data 5 is divided into learning data and test data. Specifically, the division flag according to the individual number shown in FIG. 3(b) is applied to the time-series operation data 5 shown in FIG. 3(a). As a result, the time-series operating data 5 is divided into learning data and test data. By similarly applying FIG. 3(b) to the maintenance history data 6, the maintenance history data 6 can also be divided into learning data and test data. The division method is not particularly limited, but it is preferable to divide it so that there is no bias in the number of failure occurrences between learning and testing.

Next, the time-series operating data selection unit 13 will be explained. The time-series operating data selection unit 13 includes a feature quantity calculation unit 14, a failure correlation calculation unit 15, and a selected time-series operating data number determining unit 16. Further, the time-series operation data selection unit 13 selects time-series operation data of the sensor related to the failure from the time-series operation data 5 . Then, the time-series operation data selection unit 13 adds a division flag to the selected time-series operation data using the division flag table 41, and outputs this and the maintenance history data 6.

For example, when there are 100 pieces of time-series operation data 5 of each sensor, the time-series operation data selection unit 13 selects 10 pieces of time-series operation data of sensors related to the failure. Then, the time-series operation data selection unit 13 outputs the maintenance history data and the data set 9 and the data set 91 of the selected time-series operation data. Here, the data set 9 includes both learning data and test data, and the data set 91 includes only learning data. Descriptions of the feature value calculation section 14, the failure correlation calculation section 15, and the selected time-series operation data number determining section 16, as well as a specific method for selecting time-series operation data of sensors related to failures, will be described later.

Next, the failure probability evaluation section 10 will be explained. The failure probability evaluation section 10 includes a life modeling section 17, a life model recording section 18, and a failure probability output section 19. The lifespan modeling section 17 receives the data set 91 (learning data), generates a lifespan model 20, and records it in the lifespan model recording section 18. Furthermore, the failure probability output unit 19 receives the data set 9 (learning data, test data) as input, calls the life model 21 recorded in the life model recording unit 18, and outputs the failure probability 11.

Here, the lifespan model 20 refers to a machine learning model, a statistical model, etc. that calculates failure probability from a selected set of time-series operation data. These models take time-series operation data (operation information) as input and output the corresponding failure probability. Below is a detailed explanation of each model.

A machine learning model is a model that uses machine learning methods such as neural networks (deep learning, etc.) and random forests to directly calculate failure probabilities up to the point of input based on time-series operating data (operating information). It is. Note that the machine learning method is not particularly limited to the above method. Parameters of the trained machine learning model (for example, weighting coefficients of neurons in each layer of the neural network) are recorded in the lifespan model recording unit 18.

In addition, a statistical model is a model that generates explanatory variables from a set of time-series operation data, learns the correspondence between the failure probability and explanatory variables using probability and statistical methods based on maintenance history data, and then operates over time based on the correspondence. This is a model that calculates failure probability from data. Here, when representing the correspondence relationship parametrically as a probability distribution using a statistical model, the parameters of the probability distribution are determined by maximum likelihood estimation, Bayesian estimation, or the like. Furthermore, if a statistical model does not assume a probability distribution, a non-parametric method such as the Kaplan-Meier method can be used. Note that the probability/statistical method is not particularly limited to the above method.

Further, when learning the statistical model in the life modeling unit 17, it is desirable to convert the time-series operation data according to the failure mechanism and use it as an explanatory variable of the statistical model. For example, when heat load is the dominant deterioration factor for a target such as material deterioration, how can explanatory variables be identified? Using the Arrhenius equation shown in (Equation 3), the time-series operation data T of temperature for learning data is converted into the amount of deterioration D _i that occurs at time i that is proportional to the speed of the deterioration reaction, and the cumulative value until maintenance occurs is D=ΣDi can be used as an explanatory variable.

Here, A is a constant corresponding to the deterioration reaction rate, R _a is a gas constant, and E _a is activation energy. When expressing the correspondence between the explanatory variable D and the failure probability as a parametric probability distribution F(D), for example, the Weibull distribution F shown in (Equation 4) can be assumed.

Here, the parameters α and β are the shape parameter and scale parameter of the Weibull distribution, respectively. Here, in order to identify the unknown parameters A and E _a of the Arrhenius equation and the parameters α and β of the failure probability function F (Weibull distribution) from the cumulative value D (explanatory variable) and maintenance history data for all individuals, it is necessary to The techniques described in Patent Document 1 and Patent Document 1 can be used. Furthermore, the life span model recording unit 18 records parameters A and E _a of explanatory variables and the type of probability function used for identification (Weibull distribution in the above example) as a statistical model. Note that the function identified as the failure probability function includes, in addition to the Weibull distribution, a gamma distribution, a lognormal distribution, etc., but is not particularly limited to the above.

Furthermore, in the wind power generator 1, it is known that the square value of the average wind speed is proportional to the wind load applied to the blades and the like. Therefore, the cumulative value of the square value of the average wind speed may be used as the explanatory variable. In addition, when calculating explanatory variables, we use not only the converted values of the physical quantities described above, but also the cumulative values calculated using counting methods such as the rainflow method as explanatory variables to calculate the correspondence with the failure probability (objective variable). A statistical model may be constructed to represent the The rainflow method is one of the methods of counting the stress frequency of a member subjected to repeated fluctuating loads, and is effective when the failure mechanism is fatigue fracture.

Although some simple explanatory variables were illustrated above, it is also possible to generate explanatory variables from multiple time-series operational data. Even if the failure mechanism and physical laws for the time-series operation data 5 of each sensor are unknown, the explanatory variable D(x) is automatically learned from the combination of the time-series operation data x of the target equipment and used in the statistical model. You can also do that. This method will be described later in the description of the preferred failure probability evaluation unit 10 of this embodiment (FIG. 9).

Although specific examples of machine learning models and statistical models have been shown above as lifespan models, the learning method of lifespan models is not limited to the above.

Next, the failure probability output unit 19 inputs the maintenance history data including learning data and test data, the selected data set 9, and the life model 21 stored in the life model recording unit 18. Then, the failure probability output unit 19 outputs the correspondence relationships F ₁ and F ₂ of the failure probabilities (outputs) to the time-series operation data (inputs) of the learning data and the test data.

Here, when a machine learning model is used as the life model, failure probabilities can be directly output for the selected data set (input). Therefore, the above correspondence relationship is represented by a set of pairs of input and output values obtained for each individual.

In addition, when using a statistical model as a lifespan model, the correspondence between the maintenance history data, the selected data set (input), and the failure probability (output) is the failure probability function for a parametric model, or the failure probability function for a nonparametric model. For example, it is represented by a Kaplan-Meier curve.

Next, the feature quantity calculation unit 14, failure correlation calculation unit 15, and selected time-series operating data number determination unit 16 will be specifically explained.

First, the feature calculation section 14 will be explained. The feature calculation unit 14 receives as input the time-series operation data 5 and maintenance history data 6 of each sensor. Then, the feature amount calculation unit 14 calculates the feature amounts for these and outputs survival analysis data 7. The method of calculating the feature amount is not particularly limited, but it is preferable to use the cumulative value of time-series operation data of each sensor.
During use of the object or its constituent parts, damage may accumulate and failure may occur. In the case of such a failure, for example, a wear-out failure, it is considered that the cumulative value of time-series operation data has a correlation with the failure. Here, no limitations are placed on the cumulative value calculation method, but since damage accumulation is an irreversible process, it is desirable that the cumulative value also be a monotonically increasing value. Therefore, when the time-series operating data 5 is negative, it is desirable to convert it using a function that does not take a negative value, such as an absolute value function or a soft plus function, and then calculate the accumulation. Note that it is preferable that the feature calculation unit 14 specify a division flag for the input data using the division flag table 41 and include this in the survival analysis data 7.

With the above, the cumulative value can be made to increase monotonically. Further, depending on the failure mechanism, as explained in the life modeling section, even higher precision can be achieved by applying conversion or counting methods to the time-series operation data itself and then taking cumulative values. For example, physical knowledge regarding failure mechanisms can be incorporated by applying the Arrhenius formula to temperature data, the square value to average wind speed data, and the rainflow method to stress data. Note that in calculating the cumulative value, the above conversion and counting method, or a combination of other methods may be used.

Here, FIG. 4 shows an example of the survival analysis data 7 used in this example. As shown in FIG. 4, the survival analysis data 7 is the cumulative value of the instantaneous values of the time-series operation data 5 in FIG. 3, that is, the cumulative value of the time-series operation data of each sensor at the time of measurement. recorded. Furthermore, a survival analysis flag (survival or failure) is added to the survival analysis data 7 based on the maintenance history data 6 in accordance with a maintenance event. As survival analysis flags, a survival flag is added in the case of preliminary maintenance, and a failure flag is added in the case of corrective maintenance.

In addition, in the case of an individual that has never undergone maintenance since its first operation and is still in operation (such as Unit 1 at 〇〇 site in Figure 4), there is no record in the maintenance history database. In this case, "no maintenance" is recorded in the maintenance event column in FIG. Even in the case of "no maintenance", a survival flag is added as a survival analysis flag in the same way as in advance maintenance. As a result, survival analysis data 7 is obtained in which the cumulative value of time-series operation data of each sensor at the time of measurement for each individual is paired with a survival analysis flag (survival or failure). Note that the feature amount calculation unit 14 applies the division flag table 41 to the survival analysis data 7, and adds a division column by linking the individual (site name and its machine number) to the learning data, It can be divided into test data.

Next, the failure correlation calculation section 15 will be explained. Note that the operation in the failure correlation calculation unit 15 is performed only on the learning data. This is to select the time-series operating data based on the output value of the failure correlation calculation unit 15 (the correlation between the feature amount of the time-series operating data of each sensor and the failure) and learn the life model. In other words, the test data used for prediction accuracy verification should not be used in the failure correlation calculation unit 15. The failure correlation calculation section 15 can function as a dispersion degree calculation section that indicates the degree of dispersion of the failure probability function and calculates the degree of dispersion according to the correlation with failures in parts.

The failure correlation calculation unit 15 inputs learning data whose division flag is “learning” in the survival analysis data 7, and calculates the time-series operation of each sensor using the feature calculated by the feature calculation unit 14. Calculate the correlation between data and failures, and the correlation between operating time and failures. Specifically, the failure correlation calculation unit 15 uses the cumulative value of time-series operation data of each sensor as the feature amount. Then, the failure correlation calculation unit 15 outputs a list 8 in which the dispersion of the failure probability function with respect to the cumulative value and the dispersion of the failure probability function based on the total operating time are estimated. As shown in the explanation of the life modeling unit 17, in the data set of the survival analysis data 7, for example, only the cumulative value of the time-series operation data (average wind speed) of the sensor 1 is focused. Then, a known technique can be used to estimate a failure probability function that fits the variable of interest as an explanatory variable. Therefore, the failure correlation calculation unit 15 evaluates the failure probability function that matches the cumulative value and total operation time of the time-series operation data of each sensor of the survival analysis data 7 through the same process, and calculates the number of time-series operation data. (N)+1 failure probability functions can be obtained. Once the failure probability function is determined, its variation can be easily quantified using, for example, the coefficient of variation. For example, the mean E(x) and variance V(x) of the random variable x according to the failure probability function of the two-variable Weibull distribution shown in (Equation 4) can be calculated using the following (Equation 5) and (Equation 6).

From the above, the variation coefficient C _v becomes a function only of the shape parameter α as shown in (Equation 7) below.

Here, FIG. 5 is a diagram showing the failure probability function and its coefficient of variation for the cumulative value of time-series operating data in this embodiment. Specifically, FIG. 5(a) shows a failure probability function 22 that matches the time-series operational data cumulative value (average wind speed cumulative value) of the sensor 1 related to the failure. Further, FIG. 5(b) shows a failure probability function 23 that is adapted to the cumulative value of time-series operating data of sensor N that is not related to failure. FIG. 5C shows the failure probability function 24 based on operating time. Here, as shown in FIG. 5(a), in the case of the wind power generator 1, the failure probability function of the parts (for example, the speed increaser and blades) is the cumulative value of the time-series operation data (average wind speed) of the sensor 1. The probability of failure increases rapidly at locations where the value exceeds a certain value. This is because the average wind speed is a physical quantity that correlates with the wind load applied to the wind power generator. In addition, Fig. 5 (a) shows that the variation in the failure probability function becomes smaller when the average wind speed is used as the standard. show.

As shown in FIG. 5(b), there is no correlation between the magnitude of the cumulative value and the failure probability (it is smaller). This is because the time-series operation data of sensor N is not associated with failure or has a smaller correlation. Therefore, even if the cumulative value of the time-series operation data of sensor N increases, the failure probability function does not necessarily increase rapidly, and the dispersion of the failure probability function increases.

In addition, as shown in FIG. 5(c), the operating time-based failure probability function 24 is identified based on the cumulative value of operating hours (corresponding to the number of operating days in FIG. 4) without using time-series operating data. It is.
Here, in order to utilize time-series operation information to improve prediction accuracy over time-based failure probability evaluation, it is necessary to satisfy the following conditions. In other words, it is necessary that the variation in the failure probability function with respect to the cumulative value of time-series operation data of the sensor related to failure is at least smaller than the variation in the failure probability function with respect to the total operating time. Moreover, list 8 is shown in FIG. 5(d). This is the variation coefficient of the failure probability function based on the cumulative value of time-series operating data of each sensor and the total operating time, which is output from the failure correlation calculation unit 15 to the selected time-series operating data number determining unit 16. It is quantified as . Note that FIGS. 5(a) to 5(d) can be displayed on the display unit 12.

Next, the selected time-series operating data number determination unit 16 will be explained. The selected time-series operation data number determining unit 16 includes the list 8, the maintenance history data 6, the correspondence relationship F ₁ between the failure probability (output) and the time-series operation data (input) of the learning data, and the time-series operation data of the test data. The correspondence relationship F ₂ between failure probability (output) and (input) is input. Then, the selected time-series operating data number determining unit 16 outputs the maintenance history data 6 used for failure probability evaluation and a data set 9 of N _s time-series operating data. Note that the selected time-series operating data number determination unit 16 can function as an operating data selection unit that selects operating data.

Next, a procedure for determining the number N _s of selected time-series operating data using each of the above-described configurations will be explained using the flowchart shown in FIG. 6. First, in step S50, the time-series operational data selection unit 13 selects N _s pieces of high-order time-series operational data with small variations in failure probability functions. This is done by deleting the columns of time-series operation data of each sensor shown in Figure 3 other than the selected time-series operation data of the top sensors with the smallest coefficient of variation, and only selecting the time-series operation data of the top sensors with the smallest variation. means to select. An example of the maximum value N _max of N _s is the total number N of time-series operating data.

Here, in order to shorten the calculation time, it is also effective to set the maximum value of N _s to the total number of time-series operation data corresponding to the feature quantity whose failure probability function has a smaller variation than the operation time.
Note that the minimum value is 1. The range of N _s can be initially varied from 1 to N _max at logarithmic intervals, and a rough search is performed to find a range that improves the predictive performance index described later, and then a detailed search is performed within that range in increments of 1. good.

Further, in step S51, the failure probability evaluation unit 10 inputs the selected time-series operation data and maintenance history data 6, and obtains the correspondence between the failure probability (output) and the time-series operation data (input). As a result, the failure probability evaluation unit 10 identifies the learning data correspondence relationship F ₁ for the learning data, and the test data correspondence relationship F ₂ for the test data.

Further, in step S52, the failure probability evaluation unit 10 calculates a predicted performance index of the life model from the correspondence relationship F ₁ of the learning data and the correspondence relationship F ₂ of the test data. When a machine learning model is used as the life model, the above correspondence becomes a set of pairs of failure probabilities (outputs) for time-series operating data (inputs). In this case, the degree of dissimilarity between sets may be calculated and used as an evaluation index. Examples of the degree of dissimilarity include Minkowski distance (eg, Euclidean distance and Manhattan distance), Mahalanobis distance, and cosine similarity multiplied by a negative value.

In addition, when a statistical model is used as the life model, the above correspondence can be obtained as a failure probability function or Kaplan-Meier curve that defines the input-output relationship between time-series operating data (input) and failure probability (output). It will be done. As an evaluation index, a failure probability function or a deviation rate R between Kaplan-Meier curves can be used. The deviation rate R can be calculated using (Equation 8). Note that this deviation rate R is an example of a degree of dispersion indicating a correlation.

Even if the correspondence relationship cannot be obtained as a continuous function, the deviation rate R can be calculated by calculating the difference in failure probability at each point. Furthermore, when the failure probability function is obtained as a continuous probability distribution function, the Kullback-Leibler divergence D _KL (F ₁ ||F ₂ ) can be used to quantify the difference between two distribution functions. (Equation 9) shows a calculation formula for the Kullback-Leibler divergence in the case of a continuous probability distribution as an example.

Furthermore, the performance index of the lifespan model is not limited to the above-mentioned index. For example, the difference in the variation itself in the correspondence relationship of the failure probability (output) to the time-series operating data (input) (for example, the coefficient of variation of the failure probability function _F1 of the learning data and the failure probability function _F2 of the test data) (such as a difference in coefficient of variation) can also be used.

Further, steps S50 to S52 are repeated by changing the number N _s of selected time-series operating data, and in step S53, the time-series operating data selection unit 13 determines the number N _s of selected time-series operating data for which the evaluation index is the best. .

The effects of this embodiment will be explained below using virtual data as an example. To simulate a mechanical system that obtains time-series operating data from a large number of sensors, the number of time-series operating data measured by virtual sensors is set to 100. Moreover, among them, the time-series operation data of 10 sensors are damage sensors related to the failure, and 90 are the time-series operation data of dummy sensors not related to the failure. Failures in the virtual data are generated according to the Weibull distribution of cumulative damage expressed as a linear sum of ten time-series operating data of the damage sensor.

As the life model, the damage model D(x) obtained by the damage model generation/updating unit 303 in the failure probability evaluation unit 10, which will be described later with reference to FIG. 9, is used, which is desirable for this embodiment. Furthermore, the failure probability evaluation unit identifies the Weibull distribution as the correspondence relationship between failure probabilities (output) with respect to time-series operating data (input) for learning and testing.

Here, FIG. 7 is a graph showing the relationship between the deviation rate R and the number of selected time-series operating data N _s in this embodiment. In FIG. 7, the deviation rate R is plotted on the Y-axis, and the number of selected time-series operating data N _s is plotted on the X-axis. Further, the broken line in FIG. 7 is the deviation rate based on the total operating time, and the thick solid line indicates the deviation rate of 0. The closer the deviation rate is to 0, the better the prediction performance is, but in FIG. 7, the deviation rate is the smallest when the number of selected time-series operating data is 10. Therefore, 10 can be selected as the number N _a of selected time-series operating data. This means that the deviation rate R is lower and the prediction performance is better when time-series operating data is selected (N _w =10) than when time-series operating data is not selected (N _s =100). This shows that the results are improved.

Further, FIG. 8 is a graph showing the relationship between failure probability and cumulative damage in this example.
Among them, FIG. 8A shows the failure probability evaluation result 27 when the time-series operational data selection unit 13 selects N _s =10 pieces of time-series operational data with small variations in the failure probability function. Further, FIG. 8(b) shows the failure probability evaluation results for the graph 28 when no selection is made, that is, N _s =100. In FIGS. 8(a) and 8(b), the failure probability function F ₁ (solid line A) during learning and the failure probability function F ₂ (broken line B) during testing can be compared. In Fig. 5(a), the deviation between the solid line A and the broken line B is small, and highly accurate prediction is possible, whereas in Fig. 5(b), the deviation between the solid line A and the broken line B is large, and the prediction performance deteriorates. ing. This shows that selecting time-series operational data related to failures prevents overfitting to learning data and has the effect of improving prediction accuracy (generalization performance) for unknown data. Note that FIGS. 8(a) and 8(b) can be displayed on the display unit 12. As described above, in this embodiment, the selected time-series operation data number determining unit 16 determines the difference between the correspondence of operation history data during learning (for learning) and the correspondence between operation history data during testing (for testing). The operation history data with the minimum rate (degree of deviation), that is, the minimum degree of deviation, is selected. As mentioned above, this correspondence relationship is a set of a predetermined number of pairs of time-series operational data and failure probabilities that are highly correlated with failures among the time-series operational data, or a parametric or non-parametric failure probability function. shown.

Next, a modification of the failure probability evaluation unit 10 in this embodiment different from that shown in FIG. 1 will be described.
FIG. 9 is a system configuration diagram including a modification of the failure probability evaluation section 10 in this embodiment. In this modification, the time-series operation data selection unit 13 selects time-series operation data of sensors related to wear-out failures. Therefore, the failure probability evaluation unit 10 automatically generates an explanatory variable D(x _t ) that appropriately expresses the cumulative damage to the equipment, and builds a statistical model for the explanatory variable to achieve higher effectiveness. can be expected.

Details of the failure probability evaluation unit 10 of this modification are shown below. The failure probability evaluation section 10 includes a life modeling section 17, a life model recording section 18, and a failure probability output section 19. The life modeling section 17 includes an explanatory variable generation/updating section 30 and a failure probability function identification section 33.

First, the life modeling section 17 in this modification will be explained. Here, the life modeling section 17 includes an explanatory variable generation/updating section 30 and a failure probability function identification section 33, and outputs an explanatory variable D(x). Note that the function for this can be realized by the technology described in Patent Document 1. Further, the explanatory variable D(X _t ) is expressed as a function of the time-series operation data X _t by (Equation 10).

Here, d(x) is the damage to equipment per unit time, and x is an operating data vector representing a time-series operating data set at a certain moment. In this modification, since a wear-out failure is targeted, the time integral D(x) of damage d(x) per unit time is used as an explanatory variable that causes a device to fail. Further, in this modification, the shape of the formula for calculating the explanatory variable D(x) is not particularly limited. For example, as shown in (Equation 11), the explanatory variable D(x) can be expressed in the simplest form as a linear combination of the selected N _s time-series operation data, and the optimization calculation can be performed at a relatively low calculation cost. It's over.

Here, C = (c1, c2, ... cNs) is a coefficient vector representing the weighting of each time series operation data, and x = (x _{1, x 2, ..., xNs) is the coefficient vector representing the weighting of each time series operation data, and x = (x 1,} x _2, ..., xNs) is the coefficient vector representing the weighting _of each time series operation data. This is the moment-to-moment value of operating data.

Next, the failure probability function identification section 33 will be explained. The processing of the failure probability function identification section 33, that is, the identification method of the failure probability function, is well-known like the failure probability output section 19 and the failure correlation calculation section 15. When generating the explanatory variable D(x), the explanatory variable generating/updating section 30 minimizes variations in the failure probability function while repeatedly calling the failure probability function identification section 33.

Next, the explanatory variable generation/update section 30 will be explained. The explanatory variable generation/updating unit 30 receives the time-series operation data and maintenance history data 6 used for failure probability evaluation as input, and outputs an explanatory variable D(x) and standard survival analysis data 32. Here, the explanatory variable-based survival analysis data 32 is the survival analysis data 7 to which a column of explanatory variables D(x) is added. In the survival analysis data 32, a failure probability function is obtained by identifying the failure probability function described above. Therefore, variations in the failure probability function, such as the coefficient of variation, can be easily evaluated.

Furthermore, the explanatory variable generation/updating unit 30 automatically searches for an explanatory variable that takes into account time-series operating data that minimizes the variation 31 of the failure probability function. The explanatory variables obtained as a result are reflected in survival analysis data 7 to generate survival analysis data 32 based on the damage model. The search for explanatory variables can be reduced to an optimization problem in which the objective function is varied and the explanatory variable D(x) is the parameter.

Furthermore, if the failure mechanism is known to some extent, a method may be adopted in which only the shape of the equation is defined in advance by the user according to the failure mechanism, and the coefficients thereof are searched. The conversion method and counting method as shown in the description of the life modeling section 17 and the feature value calculation section 14, and their combinations can be used.

The optimization calculation method for obtaining the coefficient vector representing the weighting of each time-series operational data is not particularly limited. However, since the objective function may be non-convex, it is desirable to use metaheuristics such as genetic algorithms or particle swarm optimization.

The explanatory variable D(x) learned as described above is stored in the life model recording section 18, and then called as appropriate by the failure probability output section 19 and used for calculating the failure probability.

Next, the display section 12 will be explained. The display unit 12 displays the failure probability up to the present time from the failure probability output unit 19, the variation in the failure probability function of the cumulative value of time-series operation data of each sensor, and the operation A list 8 that estimates the dispersion of the time-based failure probability function is input. The display unit 12 then displays these. The display unit 12 is specifically composed of a computer equipped with a screen drawing program and a display device, but the computer used here may have different functions from those of the arithmetic unit described above.

Next, the display on the display unit 12 will be explained. FIG. 10 is a diagram showing an example of the graphical user interface (GUI 40) on the display unit 12 in this embodiment. The illustrated GUI 40 shows the correlation with the failure, the sensor number, the sensor name, and the coefficient of variation for time-series operation data that are highly relevant to the failure. Here, the correlation with the failure indicates the degree of association with the failure of the device, and the higher the rank, the higher the correlation. Further, the sensor number and sensor name indicate the sensor that measured the failure. Furthermore, the coefficient of variation indicates the dispersion of the failure probability function. Therefore, in FIG. 10, the sensor numbers, sensor names, and coefficients of variation of the higher ranks (1st to 5th) that are highly relevant to failures from the time-series operation data are displayed. In this display, the time-series performance data selected by the time-series performance data selection unit 13 is displayed on the screen of the display unit 12. As a result, the user can quantitatively grasp information such as what kind of factor is the cause of component failure and how much the factor is related to the failure compared to the total operating time standard. Therefore, in this embodiment, it is possible to provide useful information not only for the maintenance and operation of the wind power generator 1 but also for the design and development of the wind power generator.

In Example 1, the failure probability for the failure that has occurred was evaluated, but in Example 2, the future failure probability is predicted and evaluated. FIG. 11 is a system configuration diagram in the second embodiment. In this embodiment, the failure probability evaluation unit 10 predicts and evaluates future failure probabilities. The difference from the failure probability evaluation unit 10 shown in FIG. 9 in the first embodiment is that it includes an operating status prediction unit 34. The operating status prediction unit 34 receives the time-series operating data 5 and the learned life model 21 as input, and outputs an explanatory variable D(x) at a future point in time.

The method for calculating future failure probability is shown below. For the individuals alive at the present moment (t ₀ ), the probability of failure after an arbitrary period of time is calculated separately. First, the predicted value 35 of the explanatory variable D(x) after an arbitrary time (Δt) has elapsed is predicted in advance by the operating status prediction unit 34. At this time, it is necessary to make predictions for each individual individually. The simplest method is to make an assumption that the average value of the values of the time-series operation data 5 recorded up to the present time will continue in the future. When using seasonally dependent operational data such as wind conditions and temperature, it is desirable to make estimates with reference to seasonal trends and forecasts from weather forecasting organizations.

Alternatively, several future operation scenarios may be assumed, and time-series operation data may be arbitrarily generated for each of them. In addition to the above, time series prediction may be performed using a state space model such as a Kalman filter. Note that the specific prediction method is not limited to the above.

By applying the estimated value of the estimated or generated time-series operation data to the life model 21 recorded in the life model recording unit 18, it is possible to calculate the explanatory variable value 35 D(t ₀ +Δt) at a future point in time. . Based on this and the current explanatory variable D(t ₀ ), the probability P that the currently operating individual will fail before an arbitrary time (Δt) has elapsed is determined by the failure probability output unit 19 as follows: (Equation 12) It can be calculated as a conditional probability according to

Here, F(D) is the correspondence relationship of failure probability (output) to time-series operating data (input) identified as a failure probability function. The failure probability is the expected value of the number of failure events occurring after an arbitrary time (Δt) has elapsed. The longer Δt is, the larger this value will be, but it is usually desirable to set Δt taking into account the periodic inspection interval of the mechanical system. .It is unlikely that the value will be close to 0. However, it is possible that the total value of failure probabilities for all individuals exceeds 1.0. This total value is the expected value of the number of failure event occurrences for the entire target individual.

Therefore, by reflecting the total value of failure probabilities in, for example, the parts inventory management system 36, it is possible to optimize the inventory status of replacement parts. Alternatively, when the operation planning system 37 is connected to the wind power generator system, a method of changing the operation plan using failure probability may be adopted. For example, if the probability of failure is higher than expected until the periodic inspection scheduled for the future, it is possible to extend the life of the parts by changing the operation plan to actively stop the wind turbine or reduce output. It becomes possible. Due to this change in the operation plan, the future operation status will also change, and the future cumulative damage will also change, so in this case, the operation plan should be reflected in the calculation of future cumulative damage in the operation status prediction unit. is desirable. With such a configuration, the user can easily confirm the relationship between changes in the operation plan and changes in failure probability.

The calculation method in this example and the current and future failure probabilities calculated using this method can be used for management (inventory management and operation planning) of mechanical systems and facilities such as factories and plants other than wind power generator systems. Can be applied. An example of this application will be explained in Example 3.

Embodiment 3 applies the failure probability at a future point in time calculated in Embodiment 2 to machinery insurance. As an example, a case will be explained in which a machinery insurance premium rate is determined using failure probability. Specifically, before the start of insurance operation, this embodiment is applied to insured equipment to construct a life prediction model. We will design machinery insurance rates based on this lifespan prediction model. In order to reduce calculation costs, it is not necessary to generate and update the model during the insurance period (usually one year). At the time of insurance renewal (rate calculation), the lifespan prediction model is updated based on newly obtained time-series operation data and maintenance history data, and the model is updated. Furthermore, the present invention may be applied not only to machinery insurance but also to insurance that covers fire, corrosion due to aging, rust, etc. in the category of failure.

In this way, when applied to insurance, the process may be executed in the cloud system shown in FIG. 12. FIG. 12 is a system configuration diagram in the third embodiment. In FIG. 12, a failure probability evaluation system 100 excluding the display unit 12 is connected to an insurance company system 101 and an asset owner system 102 via a network 1000. The insurance company system 101 and the asset owner system 102 have a network such as an intranet, and each information processing device (terminal or server) can access the failure probability evaluation system 100. Furthermore, like the terminal 103, it may be connected to the network 1000 without going through an intranet or the like. It is assumed that the display unit 12 is held by each of these terminals. Note that even when performing failure probability evaluation on a cloud system as described above, it is not always necessary to generate and update the damage model, and it may be performed as appropriate. Note that each of these devices can be realized by a so-called computer.

Each of the failure probability evaluation systems 100 of Examples 1 to 3 can be realized by a computer.
An example of this implementation is shown in FIG. FIG. 13 is a hardware configuration diagram of the failure probability evaluation system 100 in the fourth embodiment. As shown in FIG. 13, the failure probability evaluation system 100 includes a processing device 111, a memory 112, a network interface 113, and a secondary storage device 114, which are connected to each other via a communication path such as a bus.

First, the processing device 111 is a so-called processor such as a CPU, and executes processing according to the failure probability evaluation program 115 stored in the secondary storage device 114. This process is the process of each part shown in Examples 1 to 3. Further, the memory 112 stores the failure probability evaluation program 115 used for processing in the processing device 111 and the above-mentioned data.

Furthermore, the secondary storage device 114 stores a failure probability evaluation program 115 and each of the above-mentioned data, that is, a time-series operation database 2 and a maintenance history database 3. The secondary storage device 114 may be realized by various storage media such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), or a memory card. Furthermore, the failure probability evaluation system 100 may be implemented in a separate device, such as a file server.

1...Wind generator, 2...Time series operation database, 3...Maintenance history database, 4...Worker in charge, 5...Time series operation data, 6...Maintenance history data, 12...Display section, 13...Time series operation data selection Parts, 14...Feature value calculation unit, 15...Failure correlation calculation unit, 16...Selected time series operation data number determination unit, 17...Life modeling unit, 18...Life span model recording unit, 19...Failure probability output unit, 20...Life span Model, 100... Failure probability evaluation system, 101... Insurance company system, 102... Asset owner system, 103... Terminal, 1000... Network

Claims

In a failure probability evaluation system for evaluating the probability of failure in parts constituting a mechanical system,
a maintenance history database that stores maintenance history data of the mechanical system;
an operation database that stores a plurality of operation data indicating the operation status of the component;
Based on the operation data and maintenance history data, the degree of dispersion of a failure probability function for calculating the failure probability for each of the operation data is indicated, and a degree of dispersion is calculated according to the correlation with failures in the parts. a dispersion calculation unit;
an operation data selection unit that selects operation data according to the calculated dispersion degree;
A failure probability evaluation system that realizes evaluation of the failure probability by preferentially using the selected operation data.
The failure probability evaluation system according to claim 1,
Furthermore, a failure probability evaluation system includes a failure probability evaluation unit that calculates a failure probability using the maintenance history data and the selected operation data.
The failure probability evaluation system according to claim 2,
The failure probability evaluation unit includes a failure probability function identification unit that calculates a failure probability function of the mechanical system through statistical processing based on the maintenance history data;
an explanatory variable generation/updating unit having a function of generating an explanatory variable of a failure probability function that minimizes variation in the failure probability function using the operation data;
The failure probability function identification unit is a failure probability evaluation system that provides a failure probability function that minimizes variation in life.
The failure probability evaluation system according to claim 1,
The operation data selection unit is a failure probability evaluation system that selects operation history data with a minimum degree of deviation indicating the degree of deviation between the correspondence between the learning operation history data and the test operation history data.
The failure probability evaluation system according to claim 4,
The correspondence relationship is a failure probability evaluation system in which the correspondence relationship is expressed by a set of pairs of a predetermined number of upper-order operation data and failure probabilities that are highly correlated with failures among the operation data, or by a parametric or non-parametric failure probability function.
In a failure probability evaluation method for evaluating the probability of failure in parts constituting a mechanical system,
storing maintenance history data of the mechanical system in a maintenance history database;
storing a plurality of operation data indicating the operation status of the parts in an operation database;
Based on the operation data and maintenance history data, the degree of dispersion of a failure probability function for calculating the failure probability for each of the operation data is indicated, and a degree of dispersion is calculated according to the correlation with failures in the parts. ,
Selecting operation data according to the calculated dispersion degree,
A failure probability evaluation method that implements evaluation of the failure probability by preferentially using the selected operation data.
In the failure probability evaluation method according to claim 6,
Furthermore, a failure probability evaluation method that calculates a failure probability using the maintenance history data and the selected operation data.
In the failure probability evaluation method according to claim 7,
Calculating a failure probability function of the mechanical system by statistical processing based on the maintenance history data,
Using the operating data, generating an explanatory variable for the failure probability function that minimizes the variation in the failure probability function,
A failure probability evaluation method that provides a failure probability function that minimizes life dispersion.
In the failure probability evaluation method according to claim 6,
A failure probability evaluation method that selects operation history data with the minimum degree of deviation indicating the degree of deviation between the correspondence relationship of operation history data for learning and the correspondence relationship of operation history data for testing.
The failure probability evaluation method according to claim 9,
The correspondence relationship is a failure probability evaluation method in which the correspondence relationship is expressed by a set of pairs of a predetermined number of upper-order operation data and failure probabilities that are highly correlated with failures among the operation data, or by a parametric or non-parametric failure probability function.