JP7803722B2 - Failure sign monitoring system and failure sign monitoring method - Google Patents

Description

This invention relates to a failure sign monitoring system and failure sign monitoring method that monitor signs of failure before they occur, and more particularly to a failure sign monitoring system and failure sign monitoring method that monitor various types of plants.

Power plants at power plants use various types of equipment, such as steam turbines and boilers, to generate electricity. These power plants monitor the installed equipment for signs of failure (see, for example, Patent Document 1). According to the monitoring system described in this document, System Invariant Analysis Technology (SIAT) is used to monitor for signs of failure. The invariant analysis technology utilizes a network fault response engine. The network fault response engine monitors the entire system by detecting disruptions in the relationships between observation points.

This invariant analysis technology is applied to plant monitoring. That is, there is a strong correlation between the measurements of each sensor in the plant (hereinafter referred to as "sensor values"). Conventional monitoring systems automatically analyze sensor values collected from observation points, comprehensively discover and model correlations (invariants), which are invariant relationships between sensor values between two points, and create a correlation model (hereinafter referred to as "monitoring model") for monitoring equipment, etc. The monitoring system then uses the monitoring model to examine the correlation between, for example, a sensor installed on a pipe in a power plant and a sensor installed on a steam turbine for any "unusual" behavior, i.e., any disruption in the relationship. The monitoring system can then detect abnormalities in the power plant from "unusual" behavior, making it possible to monitor the entire plant in real time.

Japanese Patent Application Laid-Open No. 2017-021702

However, the conventional monitoring system mentioned above has the following issue. Specifically, this monitoring system has an automatic update function that automatically updates the monitoring model on a regular basis. However, using the automatic update function raises concerns about "missed detections" and "false positive detections." For example, there is a missed detection (Example 1) as shown in Figure 6. If the monitoring system evaluates the period "third week of month xx of year xx" using a monitoring model based on sensor values from "second week of month xx of year xx," it is possible to detect "Trouble I." Note that in Figure 6, the sensor values are represented by two curves, upper and lower. The lower curve g1 shows the sensor value when there is no trouble and the upper curve g2 shows the sensor value when a trouble, or failure, occurs.

However, if the monitoring system evaluates the period "week 4 of month xx of year 20xx" using a monitoring model based on sensor values from "week 3 of month xx of year 20xx," it will learn the sensor value fluctuations for "Trouble I" in week 3 as normal, and may not be able to detect "Trouble II," which involves similar sensor value fluctuations.

There is also the possibility of false positives (Example 2) as shown in Figure 7. For example, even if the sensor value from a given sensor is constant due to maintenance in "week 3 of month xx of year 20xx," if the monitoring system creates a model and evaluates "week 4 of month xx of year 20xx," it will learn that the constant sensor value from the sensor is normal behavior. For this reason, the monitoring system may detect a status change in "week 4 of month xx of year 20xx" even when nothing is happening.

The object of this invention is to solve the above-mentioned problems and provide a failure sign monitoring system and failure sign monitoring method that can prevent missed detections and false positives when monitoring for signs of plant failures using a monitoring model created with an automatic update function.

In order to solve the above-mentioned problems, the invention of claim 1 is a plant monitoring system that is used for a plant that is equipped with various types of equipment and various sensors that measure the states of these equipment, monitors the plant using a monitoring model created based on correlations between the sensors, and periodically updates this monitoring model, and is characterized by comprising: processing means that, before updating the monitoring model using each sensor value of the current period, compares a variance ratio, which is the ratio of the variance calculated from each sensor value of the current period to the variance calculated from each sensor value of the period prior to the current period, with a preset reference value, and, if the variance ratio is greater than the reference value, excludes sensor values whose variance ratio is greater than the reference value and updates the monitoring model.

The invention of claim 2 is a failure sign monitoring method used for a plant equipped with various devices and various sensors for measuring the states of these devices, which monitors the plant using a monitoring model created based on correlations between the sensors and periodically updates this monitoring model, characterized in that before updating the monitoring model using each sensor value for a current period, a variance ratio, which is the ratio of a variance calculated from each sensor value for the current period to a variance calculated from each sensor value for a period prior to the current period, is compared with a preset reference value, and if the variance ratio is greater than the reference value, sensor values whose variance ratio is greater than the reference value are excluded and the monitoring model is updated .

According to the inventions of claims 1 and 2 , when a monitoring model is periodically updated, if the sensor values for the current period have large variations, the monitoring model is created from each sensor value obtained by removing the sensor values with large variations or from each sensor value obtained from another period. As a result, when signs of failure in equipment installed in a plant are monitored in the next period using the updated monitoring model, a monitoring model created using sensor values with large variations is not used, thereby preventing missed detections and false detections of signs of failure and preventing a decrease in monitoring accuracy.

According to the first and second aspects of the present invention, by calculating the variance of the sensor values in the current period, it becomes possible to check whether the variations in the sensor values in the current period are large.

According to the inventions of claims 1 and 2 , by calculating the variance ratio between the variance of the sensor values in the previous period and the variance of the sensor values in the current period, it is possible to check using a simple calculation method whether the variance of the sensor values in the current period is large.

1 is a configuration diagram showing a plant monitoring system according to a first embodiment of the present invention; FIG. 10 is a diagram illustrating an example of sensor data. FIG. 10 is a diagram illustrating an example of time-series sensor values. 10 is a flowchart illustrating an example of a determination process. FIG. 10 is a diagram illustrating a part of the calculation for calculating variance from sensor values. FIG. 10 is an explanatory diagram illustrating an example of a detection failure. FIG. 10 is an explanatory diagram illustrating an example of erroneous detection.

Next, each embodiment of the present invention will be described in detail using the drawings.

(Embodiment 1)
A plant monitoring system according to this embodiment is shown in Fig. 1. The plant monitoring system of Fig. 1 is used in a power plant where a power plant 10 is installed, and mainly comprises a sensor monitoring device 20 and a model monitoring device 40. The sensor monitoring device 20 is connected to the model monitoring device 40 via an in-house communication network 30 such as an intranet so that data communication with the model monitoring device 40 is possible.

The power plant 10 generates electricity using nuclear or thermal power, and this embodiment illustrates an example in which nuclear power is used. The power plant 10 uses a large number of devices, such as a nuclear reactor, turbines, generators, pumps, and piping, although these devices are not shown. The power plant 10 generates electricity using these devices. The power plant 10 also has various sensors, such as _A1 sensor to _Am sensor, etc., which are also not shown in the drawings, to check the status of each device.
Sensors B ₁ to B _n are installed. Each sensor outputs a sensor value a ₁ to a _m
...
b ₁ to b _n
is sent to the sensor monitoring device 20.

The sensor monitoring device 20 receives sensor values _a1 to _am , ..., _b1 to _bn from each sensor installed in the power plant 10. When the sensor monitoring device 20 receives the sensor values _a1 to _am , ..., _b1 to _bn from each sensor, it creates sensor data, for example, as shown in FIG. 2. The sensor data indicates the sensor and records the installation point where the sensor is installed corresponding to sensor identification information for identifying the sensor. The sensor data also records the sensor value and measurement date and time of the sensor corresponding to the sensor identification information. FIG. 2 shows an example in which the sensor values _a1 to _am are from sensors _A1 to _Am installed in the condenser of the power plant 10, and the sensor values _b1 to _bn are from sensors _B1 to _Bn installed in the feedwater pump system of the power plant 10. The configuration will be described below based on this example.

The sensor monitoring device 20 transmits the created sensor data to the model monitoring device 40 via the in-house communication network 30 each time the transmission time elapses. The transmission time is set in the sensor monitoring device 20 as needed, such as in seconds, minutes, hours, or days.

The model monitoring device 40 receives sensor data from the sensor monitoring device 20 via the in-house communication network 30. The model monitoring device 40 uses the received sensor data to monitor the power plant 10. To this end, the model monitoring device 40 is equipped with a communication control unit 41, a data server 42, a management server 43, and clients 44 ₁ to 44 _k . The communication control unit 41, data server 42, management server 43, and clients 44 ₁ to 44 _k of the model monitoring device 40 are connected by a LAN (Local Area Network) or the like to enable data transmission and reception.

The communication control unit 41 performs communication control for connecting the data server 42, the management server 43, and the clients 44 ₁ to 44 _k to the in-house communication network 30. For example, when the communication control unit 41 receives sensor data from the sensor monitoring device 20 via the in-house communication network 30, it sends this sensor data to the data server 42.

The data server 42 is a storage device that stores data related to the power generation plant 10. For example, when the data server 42 receives sensor data from the sensor monitoring device 20 via the in-house communication network 30 and the communication control unit 41, it stores this sensor data. In addition, when the data server 42 receives a data transmission request from the management server 43, it extracts the corresponding sensor data and sends it to the management server 43.

The clients 44 ₁ to 44 _k are computers operated by the person in charge of operating the power plant 10, and are used to operate the power plant 10. Various instructions and the like required for operating the power plant 10 are input into the clients 44 ₁ to 44 _k by the person in charge. For example, a model creation instruction is input into the clients 44 ₁ to 44 _k in order to create each monitored object model of the power plant 10 as needed. When a model creation instruction is input, the clients 44 ₁ to 44 _k send the model creation instruction to the management server 43.

Furthermore, the clients 44 ₁ to 44 _k output an alarm when they receive various data, such as alarm data described below, from the management server 43. Furthermore, when the clients 44 ₁ to 44 _k receive sensor list data from the management server 43, they display this data, etc.

The management server 43 is a computer that performs various processes to monitor the power plant 10 and each piece of equipment. First, the management server 43 creates a monitoring model of the power plant 10. Next, the management server 43 monitors the power plant 10 and equipment based on the created monitoring model. Below, we will explain in order the creation of a monitoring model of the power plant 10 and the monitoring of the power plant 10, etc. using the monitoring model.

The management server 43 creates a monitoring model during normal operation in order to monitor the power plant 10. The monitoring model is created by using invariant analysis technology to derive correlations between sensors based on time-series data from each sensor attached to the power plant 10. The monitoring model during normal operation is created by the management server 43 based on sensor values when the power plant 10 was operating normally. In other words, since the state of the power plant 10 is stable during constant-thermal operation, the distribution of sensor values is also stable. The management server 43 creates the monitoring model based on this state.

Furthermore, in this embodiment, the management server 43 automatically updates the surveillance model at predetermined intervals, for example, once a week. The predetermined interval for automatically updating the surveillance model can be any interval, such as several days or several weeks. The management server 43 sends a data transmission request to the data server 42 at a preset timing. The management server 43 also sends a data transmission request to the data server 42 when it receives a model creation instruction manually input by a person in charge from one of the clients 44 ₁ to 44 _k .

Thereafter, the management server 43 receives each sensor data from the data server 42. Each sensor data received by the management server 43 is data during normal operation of the power plant 10, and is, for example, as shown in FIG. 2. Then, the management server 43 creates a time-series sensor value list using the sensor values obtained from each sensor data. An example of this time-series sensor value list is shown in FIG. 3. FIG. 3 shows a list of sensor values divided into predetermined periods of one week, and for the sensor value _a1 of the _A1 sensor, the time-series sensor values _a1 ( _p1 ), ..., _a1 ( _pt ) are
a ₁ (q ₁ ),..., a ₁ (q _t )
a ₁ (r ₁ ),..., a ₁ (r _t )
a ₁ (s ₁ ),..., a ₁ (s _t )
In other words, the time series sensor value for the first week of "20xx year xx month" is
a ₁ (p ₁ ),..., a ₁ (p _t )
and the time series sensor value for the second week is
a ₁ (q ₁ ),..., a ₁ (q _t )
In addition, the time series sensor value for the third week is
a ₁ (r ₁ ),..., a ₁ (r _t )
The time series sensor value for the fourth week is
a ₁ (s ₁ ),..., a ₁ (s _t )
The time series sensor value a ₁ (s _t ) for the fourth week is the last value for this week.

When the management server 43 finishes creating this time-series sensor value list, conventionally it would create a monitoring model to be used for the next week using the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) for the model learning period (week 4). In other words, conventionally, the current period, week 4, is set as the model learning period to create the monitoring model.

In contrast, this embodiment is described as follows. That is, once the management server 43 has finished creating a list of time-series sensor values, it performs a judgment process. An example of this judgment process is shown in FIG. 4. When the management server 43 receives the last sensor value _a1 ( _st ) of this week, which is the current model learning period (week 4), it starts the judgment process before updating the monitoring model. After starting the judgment process, the management server 43 calculates the variance _σi of the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) of the current model learning period (week 4) (step S1). That is, in step S1, the management server 43 checks whether the distribution of the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) of the current period (week 4, model learning period) is stable. To do this, the management server 43 calculates the average value _a1av of the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) using the following formula:

Next, as shown in FIG. 5 , the management server 43 calculates the difference between each of the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) and the average value _a1av . Thereafter, the management server 43 calculates the square of the difference between each of the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) and the average value _a1av . Next, the management server 43 adds up the squared values of the differences from the average value _a1av . Furthermore, the management server 43 divides the calculated sum by the value t to calculate the variance _σi for this week (week 4), which is the model learning period.

The value t here is the number of data t of the sensor values a ₁ (s ₁ ), . . . , a ₁ (s _t ).

When step S1 is completed, the management server 43 similarly calculates the variance σ _i-1 for the third week, which is the previous period of the current model learning period, using the sensor values a ₁ (r ₁ ), ..., a ₁ (r _t ) (step S2).

After step S2 is completed, the management server 43 performs an F-test on the distribution of the sensor values (step S3). That is, the management server 43 calculates the variance ratio F _i , which is the ratio of the variance σ _i of the model learning period (fourth week) to the variance σ _i-1 of the previous period (third week). The variance ratio F _i is a value for determining whether the model learning period (fourth week) and the previous period (third week) have equal variances. For example,

Then, the management server 43:

The variance ratio F _i is calculated by the following equation.

Then, the management server 43:

The variance ratio F _i is calculated as follows.

The management server 43 holds a predetermined judgment reference value _Fs for performing the F test. The judgment reference value _Fs is a value that serves as a judgment reference for the F test, and is calculated based on two periods when the power plant is normal, for example, the variance of the first week and the variance of the second week. The judgment reference value Fs is also a value input to the management server 43 by a person in charge, for example, operating the client ₄₄₁ , that is, the judgment reference value of the variance ratio F _i set according to the characteristics of the sensors of the power plant.

The management server 43 calculates the variance ratio F _i calculated as above as follows:

If the relationship holds, it is determined in step S3 that the variance σ _i for this week (week 4), which is the model learning period, is the same as the variance σ _i-1 for the previous week (week 3), which is the previous period. In other words, the management server 43 calculates the sensor values a ₁ (s ₁ ), ..., a ₁ (s _t ) for the model learning period.
and the sensor values a ₁ (r ₁ ), ..., a ₁ (r _t ) of the previous period.
It is determined that the variation in the

After the determination in step S3, the management server 43 creates an optimal monitoring model using the sensor values a ₁ (s ₁ ), . . . , a ₁ (s _t ) during the scheduled model learning period (step S4).

In step S4, the management server 43 creates a monitoring model using the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) for the scheduled model learning period. That is, the management server 43 checks the strength of the correlation between the sensors and creates a monitoring model using correlated sensors. At this time, the management server 43 creates in advance a monitoring model that focuses on the strength of the correlation between the sensors according to the plant state (under inspection, during startup, operating at rated thermal output, undergoing regular testing, or shut down).

On the other hand, the management server 43 calculates the variance ratio F _i calculated in step S3 as follows:

If the relationship does not hold, the management server 43 notifies, for example, the clients 44 ₁ to 44 _k of an abnormality in the sensor values (step S5). In step S5, the management server 43 creates alarm data notifying that the distribution of the sensor values is unstable as an abnormality, and sends this alarm data to the clients 44 ₁ to 44 _k .

After step S5, the management server 43 modifies or changes the model learning period as necessary by changing the model learning period or excluding from the model learning period sensor values whose variance ratio F _i is greater than the set reference value, and creates a monitoring model using the modified or changed model learning period (step S6). For example, as shown in FIG. 6 above, if the sensor value fluctuates during the model learning period due to a problem, or as shown in FIG. 7, if the sensor value becomes constant during the model learning period due to maintenance, the management server 43 detects this as a change in the variance ratio F _i . As a result, the management server 43 modifies or changes the model learning period, and therefore, in step S6, when updating the monitoring model, the management server 43 can create an optimal monitoring model equivalent to one when the distribution of sensor values is stable.

When step S4 or step S6 is completed, the management server 43 ends the judgment process. The management server 43 then performs real-time monitoring or core invariant method monitoring using the optimal monitoring model for the next week.

The above is the configuration of the failure sign monitoring system according to this embodiment. Next, we will explain the failure sign monitoring method used by this failure sign monitoring system.

Normally, the model monitoring device 40 monitors the state of the power plant 10 using a monitoring model. The management server 43 of the model monitoring device 40 then automatically updates the monitoring model at each model learning period, which is a predetermined period, for example, once a week. At this time, the management server 43 uses the judgment process shown in Figure 4 to determine whether there is any variation in the sensor values for this week, i.e., the fourth week, compared to the sensor values for the third week.

If there is no variation between the sensor values for the fourth week and the sensor values for the third week, the management server 43 creates a monitoring model using the sensor values for the fourth week and updates the monitoring model.

If there is a variation in the sensor values for the fourth week compared to the sensor values for the third week, the management server 43 corrects or changes the model learning period, creates a monitoring model using the sensor values for the corrected or changed model learning period, and updates the monitoring model. At the same time, the management server 43 creates alarm data and sends it to clients 44 ₁ to 44 _k . When clients 44 ₁ to 44 _k receive the alarm data, they notify the person in charge that the distribution of the sensor values for the model learning period is not stable.

Then, the following week, the management server 43 of the model monitoring device 40 uses the updated monitoring model to perform real-time monitoring or core invariant method monitoring of the power plant 10.

Thus, according to this embodiment, a function for performing a test based on a variance ratio (F test) on the sensor value distribution is added. In other words, a filter for the test based on a variance ratio (F test) is applied to the sensor value distribution. As a result, if the sensor value distribution is unstable, the model learning period is changed or sensor values whose variance ratio F _i is greater than the set reference value are excluded from the model learning period, making it possible to prevent "missed detections" and "false detections" when the monitoring model is automatically updated. This makes it possible to prevent a decrease in monitoring accuracy.

(Embodiment 2)
In this embodiment, the variation in each sensor value for the current period is determined as follows: In this embodiment, components that are the same as or considered to be the same as those in the first embodiment described above are given the same reference numerals, and their description will be omitted.

In this embodiment, the determination process (FIG. 4) is simplified. That is, the management server 43 calculates the variance _σi for this week (the fourth week), which is the model learning period, using the sensor values _a1 ( _s1 ), ..., _a1 ( _st ), in the same manner as in step S1 of the determination process (FIG. 4).

The management server 43 stores in advance a variance reference value that is calculated from each variance during past automatic updates and represents the normal state of the sensor values. The management server 43 then compares this variance reference value with the variance _σi for this week (week 4) to determine whether the distribution of the sensor values _a1 ( _s1 ), ..., _a1 ( _st ) during the model learning period is stable.

The management server 43 then performs step S4 or steps S5 and S6, which are the same as the judgment process (Figure 4).

Thus, according to this embodiment, by adding processing for simplified judgment based on variance to the sensor value distribution, it is possible to prevent "missed detections" and "false detections" when automatically updating the monitoring model. This makes it possible to prevent a decrease in monitoring accuracy.

10 Power plant 20 Sensor monitoring device 40 Model monitoring device 41 Communication control unit 42 Data server 43 Management server (processing means)
44 ₁ to 44 _k clients

Claims (2)

A plant monitoring system that is used for a plant that is equipped with various devices and various sensors that measure the states of the devices, monitors the plant using a monitoring model created based on correlations between the sensors, and periodically updates the monitoring model,
a processing means for comparing a variance ratio, which is a ratio between a variance calculated from each sensor value in the current period and a variance calculated from each sensor value in a period prior to the current period, with a preset reference value before updating the surveillance model using each sensor value in the current period, and, if the variance ratio is greater than the reference value, excluding sensor values whose variance ratio is greater than the reference value and updating the surveillance model;
A failure sign monitoring system comprising: A failure sign monitoring method is used for a plant that is equipped with various devices and various sensors that measure the states of the devices, the method monitors the plant using a monitoring model created based on correlations between the sensors, and periodically updates the monitoring model,
before updating the surveillance model using each sensor value for the current period, a variance ratio, which is a ratio between a variance calculated from each sensor value for the current period and a variance calculated from each sensor value for a period prior to the current period, is compared with a preset reference value, and if the variance ratio is greater than the reference value, sensor values whose variance ratio is greater than the reference value are excluded and the surveillance model is updated ;
A failure sign monitoring method characterized by: