JP7281394B2 - Abnormal diagnosis device and program - Google Patents

Description

The present invention relates to an abnormality diagnosis device and program.

In monitoring the condition of machines such as wind turbines, sensors installed in the machines are used to collect physical quantities such as temperature and acceleration of the machine at a certain frequency, and the collected data is analyzed and processed to determine the state of the machine. is monitored and judged as normal/abnormal. In recent years, machine learning technology, which is a field of artificial intelligence technology, has been used to learn (also called “training”) machine operation data collected in the past, and to estimate and judge the current state of an abnormality diagnosis method. It has been proposed (Patent Document 1 and Non-Patent Document 1 below). For example, in the abstract of Patent Literature 1 below, it is stated that ``the equipment status monitoring device, based on a normal characteristic quantity group extracted from the state quantity fluctuation data of the monitoring target equipment in a normal state, An anomaly degree calculation model creation unit that creates an anomaly degree calculation model for calculating an anomaly degree of the monitoring-time feature value group, and an anomaly degree calculation unit that calculates the anomaly degree of the monitoring-time feature value group using the anomaly degree calculation model an abnormality determination unit that determines whether or not there is an abnormality in the equipment to be monitored based on the degree of abnormality; Based on a causal matrix showing the contribution degree and the relationship between the cause of the abnormality of the equipment to be monitored and the plurality of feature values, an abnormality cause and an abnormality cause identification unit that identifies the ".

JP 2019-128704 A

Suzuki, Uchiyama, Yuda: Anomaly detection technology using data mining. Operations Research, September 2012, pp. 506-511

In the above-described abnormality diagnosis method based on machine learning, part of the measurement data from the equipment to be monitored is used as input data. Hereinafter, the measurement data used for the input of the abnormality diagnosis method will be referred to as "feature amount" or "feature amount group". In a device with a complicated structure such as a wind power generator, the types and number of measurement data are enormous. Therefore, depending on the selection state of the measurement data to be applied as the feature amount group, the result of abnormality diagnosis may vary greatly. Conventionally, a user often selects a feature amount group based on subjective judgment, and the result of abnormality diagnosis may be inaccurate.
As described above, Patent Literature 1 describes "calculating the degree of abnormality of a group of feature values during monitoring" and "calculating the degree of contribution to the degree of abnormality by each of a plurality of feature values." However, if the selection of the "feature group used for abnormality diagnosis" is inappropriate, there is a problem that the calculation results of the "abnormality degree" and "contribution degree" become inappropriate.
SUMMARY OF THE INVENTION It is an object of the present invention to provide an abnormality diagnosis apparatus and program capable of accurately determining the state of a diagnosis target apparatus.

In order to solve the above problems, the abnormality diagnosis apparatus of the present invention comprises a database for storing measurement data of a plurality of measurement items in a diagnosis target device, a failure mode selection section for selecting a failure mode to be detected, and the diagnosis target device. or a teacher data creation unit for creating teacher data for determining the presence or absence of a failure in the device to be diagnosed based on failure response measurement data that is data corresponding to the failure mode in another device; a measurement item importance calculation unit that calculates the importance of the measurement item in the failure mode based on the above, and selects a part of the measurement item as a feature amount for the failure mode based on the calculated importance A feature amount selection unit, an abnormality degree calculation unit that calculates an abnormality degree corresponding to the failure mode based on the measurement data related to the feature amount, and a diagnosis target device based on the calculated abnormality degree and a device state determination unit that determines the state.

According to the present invention, the state of the device to be diagnosed can be accurately determined.

1 is a block diagram of an abnormality diagnosis system according to a preferred first embodiment; FIG. It is a figure which shows an example of the data structure of measurement data. It is a figure which shows an example of failure data. It is a figure which shows an example of teacher data. It is a figure which shows an example of importance data. It is a figure which shows an example of the time distribution of the degree of abnormality. 4 is a flowchart of an abnormality diagnosis processing routine executed in the abnormality diagnosis device;

[First embodiment]
<Configuration of the first embodiment>
FIG. 1 is a block diagram of an abnormality diagnosis system 1 according to a preferred first embodiment.
In FIG. 1, the abnormality diagnosis system 1 includes a diagnosis target device 10, a sensor section 20, and an abnormality diagnosis device 100 (computer). In addition, in this embodiment, the diagnosis target device 10 is a wind turbine generator. The sensor unit 20 includes N (N is plural) sensors 22-1 to 22-N for measuring physical quantities such as temperature, pressure, acceleration, etc. in each part of the device 10 to be diagnosed. Then, the sensor unit 20 samples the physical quantities, which are the measurement results of the sensors 22-1 to 22-N, at predetermined sampling intervals, and supplies the results to the abnormality diagnosis device 100 as measurement data DM. The abnormality diagnosis device 100 diagnoses the state of the diagnosis target device 10 based on the measurement data DM supplied from the sensor section 20 .

The abnormality diagnosis device 100 includes general computer hardware such as a CPU (Central Processing Unit), RAM (Random Access Memory), ROM (Read Only Memory), and SSD (Solid State Drive). stores an OS (Operating System), application programs, various data, and the like. The OS and application programs are developed in RAM and executed by the CPU. In FIG. 1, inside the abnormality diagnosis device 100, functions implemented by application programs and the like are shown as blocks.

That is, the abnormality diagnosis apparatus 100 includes a measurement data acquisition unit 101, a failure data acquisition unit 102, a teacher data creation unit 103 (teacher data creation means), and a measurement item importance calculation unit 104 (measurement item importance calculation means). , feature quantity selection unit 105 (feature quantity selection means), failure mode selection unit 106 (failure mode selection means), abnormality degree calculation unit 107 (abnormality degree calculation means), and device state determination unit 108 (device state determination means) and a database 110 (database means).

The measurement data acquisition unit 101 acquires measurement data DM from the sensor unit 20 . The database 110 has one or more storage devices and stores various data in electronic file format. For example, database 110 stores the data listed below. The contents of each data will be described later.
・Measurement data DM acquired by the measurement data acquisition unit 101
・Past failure data DF
・Teacher data DT
- Importance data DQ indicating the importance of measurement data,
・ Time-series data of the degree of abnormality A, which is the result of abnormality diagnosis,
state data DC,
・Failure mode list ML

FIG. 2 is a diagram showing an example of the data structure of measurement data DM.
In the illustrated example, sampling times ts_1, ts_2, . . . ts_max are times in each sampling period. The measurement data DM includes measurement results at sampling times ts_1, ts_2, . Thus, the measurement data DM is a set of time-series data collected from each sensor 22-1 to 22-N, and can be described as a matrix as shown in FIG.

Returning to FIG. 1 , when a failure is found in the diagnosis target device 10 , the failure data acquisition unit 102 records failure data DF related to the failure in the database 110 .
FIG. 3 is a diagram showing an example of failure data DF.
As illustrated, the failure data DF includes a failure discovery date and time DFT, a failure mode DFM, and failure response measurement data DFK. Here, the failure discovery date and time DFT is data indicating the date and time when a user or a failure monitoring device (not shown) discovered a failure in the diagnosis target device 10 .

Further, the failure mode DFM is data corresponding to the specific content of the detected failure, and includes, for example, "damage to gearbox medium speed shaft pinion" and "damage to generator bearing" as shown in the figure. Further, the failure response measurement data DFK is a part of the measurement data DM described above, and is equal to the measurement data DM within a predetermined period up to the failure discovery date and time DFT.

Returning to FIG. 1, the failure mode list ML stored in the database 110 is a list listing various failure modes that can be selected as the failure mode DFM (see FIG. 3) described above. Also, the teacher data creation unit 103 creates teacher data DT based on the failure data DF described above.

FIG. 4 is a diagram showing an example of teacher data DT.
In the illustrated example, sampling times ts_a, ts_b, ts_c, ts_d, etc. are times in each sampling cycle. However, in FIG. 4, the time progresses from bottom to top, and the sampling time ts_d is equal to the failure discovery date and time DFT (see FIG. 3). Sampling time ts_c is a time earlier than ts_d by a predetermined time T1 (first predetermined time), and sampling time ts_b is a sampling time immediately before ts_c. Also, the sampling time ts_a is a time earlier than ts_b by a predetermined time T2 (second predetermined time).

The teacher data DT includes measurement results at the sampling times ts_a to ts_d for each of the measurement items P1 to PN by the sensors 22-1 to 22-N (see FIG. 1). Furthermore, the teacher data DT includes a flag called an anomaly flag DTF. In other words, the teacher data DT includes the failure response measurement data DFK included in the failure data DF (see FIG. 3) and the abnormality flag DTF. The abnormality flag DTF is "1" for the sampling times ts_c to ts_d, and is "0" for the sampling times ts_a to ts_b.

Here, the sampling time at which the abnormality flag DTF is "1" is the time at which "abnormality is highly likely to occur", and the sampling time at which the abnormality flag DTF is "0" is the time at which "abnormality is likely to occur". It is "low" time. The teacher data DT in the range where the abnormality flag DTF is "1" is called abnormal teacher data DT1, and the teacher data DT in the range where it is "0" is called normal teacher data DT2. Here, the values of the predetermined times T1 and T2 may be determined based on component knowledge related to the corresponding failure mode DFM (see FIG. 3), operation history of the device 10 to be diagnosed, and the like.

Returning to FIG. 1, the measurement item importance calculation unit 104 calculates importance data DQ based on the teaching data DT described above. Here, the importance data DQ includes N importances Q1 to QN respectively corresponding to N measurement items (measurement results of the sensors 22-1 to 22-N). There is a correlation between the N measurement items and the value of the abnormality flag DTF, and the degrees of importance Q1 to QN indicate the degree of influence each of the N measurement items has on the value of the abnormality flag DTF. . For example, "contribution to the value of the abnormality flag DTF" can be used as the degrees of importance Q1 to QN. However, the degrees of importance Q1 to QN are not limited to the "contribution", and may be other indicators such as the "contribution rate" that indicate the degree of influence on the value of the abnormality flag DTF.

The method of calculating the importance levels Q1 to QN is, for example, about the teacher data DT, using a machine learning algorithm such as a decision tree, learning is performed on the abnormal teacher data DT1 and the normal teacher data DT2, and as a result of the calculation, It is preferable to obtain the importance of each measurement item included in the failure response measurement data DFK. Moreover, it is preferable to normalize the importance levels Q1 to QN so that the sum of these values becomes a predetermined value, for example, "1".
FIG. 5 is a diagram showing an example of importance data DQ. In the illustrated example, the measurement items are listed in order of importance.

Returning to FIG. 1, the feature quantity selection unit 105 extracts M (where M<N) measurement items in descending order of importance from the importance data DQ, and determines the measurement data related to the extracted measurement items as abnormal. Selected as features for diagnosis. The number of features M may be set according to part knowledge related to the corresponding failure mode DFM (see FIG. 3), user designation, etc., but is often in the range of "2" to "10", for example. . However, when higher diagnostic accuracy is required, the number of feature values M may be increased. A plurality of feature values selected for abnormality diagnosis are called selected feature values DQS (feature values, see FIG. 5). In the example shown in FIG. 5, the number of features M is set to "3", and the measurement data related to the three measurement items "wind speed_Average", "generator bearing rear acceleration_Max", and "nacelle acceleration_Average". is selected as the selected feature quantity DQS.

In FIG. 1, the failure mode selection unit 106 selects a failure mode DFM to be diagnosed from a preset failure mode list ML (see FIG. 1) based on user's operation. Further, the degree-of-abnormality calculation unit 107 creates a normal model based on the selected feature quantity DQS (see FIG. 5) and the normal teaching data DT2 (see FIG. 4). Quantify the distance to and calculate the degree of anomaly. For example, the abnormality degree calculation unit 107 can calculate the abnormality degree A of the diagnosis target device 10 based on the following formula (1) using a statistical algorithm called the Mahalanobis-Taguchi method. In the following formula (1), x is an M-dimensional feature amount vector by the selected feature amount DQS, μ is the average value of the feature amount vector x in the normal training data DT2, and σ is the feature in the normal training data DT2. is the variance of the quantity vector x.
A=(x−μ) ² /σ ² … Formula (1)

Based on the value of the degree of abnormality A, the device state determination unit 108 determines whether the state of the diagnosis target device 10 is normal or abnormal with respect to the failure mode DFM. For example, regarding the failure mode DFM, if the degree of abnormality A is equal to or greater than a predetermined threshold value A_th, it can be determined as "abnormal", and if the degree of abnormality A is less than the threshold value A_th, it can be determined as "normal". The device state determination unit 108 stores the determination result of whether or not the state of the diagnosis target device 10 is normal in the database 110 as the state data DC (see FIG. 1) regarding the failure mode DFM. In other words, the state data DC is data indicating whether or not the diagnosis target device 10 is normal for each of the failure modes DFM.

FIG. 6 is a diagram showing an example of the time distribution of the degree of abnormality A. As shown in FIG. In FIG. 6, the horizontal axis is time, and the vertical axis is the degree of abnormality A. In FIG. In the illustrated example, there is a timing when the degree of abnormality A becomes equal to or greater than the threshold A_th, so the device state determination unit 108 determines that "the diagnosis target device 10 is abnormal" for the corresponding failure mode DFM. The processing in this device state determination unit 108 can be described in pseudo code as shown in the following equation (2).

IF (Is there a timing when the degree of abnormality A≧threshold A_th?)
THEN The diagnosis target device 10 is abnormal.
ELSE Diagnosis target device 10 is normal. … formula (2)

<Operation of the first embodiment>
Next, operation of the first embodiment will be described.
FIG. 7 is a flow chart of an abnormality diagnosis processing routine executed in the abnormality diagnosis device 100. FIG. Here, the abnormality diagnosis processing routine is classified into an offline processing routine R10 that is executed before the degree of abnormality A is calculated, and an online processing routine R20 that calculates the degree of abnormality A.

When the process proceeds to step S12 in the offline processing routine R10, the teacher data creation unit 103 creates teacher data DT corresponding to each failure mode DFM (see FIG. 3) based on the measurement data DM and the failure data DF. create. Next, when the process proceeds to step S14, the measurement item importance calculator 104 performs decision tree learning on the teacher data DT to create importance data DQ. Next, when the process proceeds to step S<b>16 , the feature amount selection unit 105 selects the selected feature amount DQS for each failure mode DFM, and stores the selection result in the database 110 .

Further, when the process proceeds to step S22 in the online processing routine R20, the failure mode selection unit 106 selects a failure mode DFM to be diagnosed based on the user's operation. Next, when the process proceeds to step S24, the degree-of-abnormality calculation unit 107 calculates the selected feature amount DQS (see FIG. 5), the normal teaching data DT2 (see FIG. 4), and the above-described formula (1). Based on this, the time-series distribution of the degree of anomaly A, that is, the degree of anomaly A at each sampling time is calculated. Next, when the process proceeds to step S26, the device state determination unit 108 determines whether or not there is an abnormality in the diagnosis target device 10 for the corresponding failure mode DFM based on the time-series distribution of the degree of abnormality A, and the determination result is state data DC (see FIG. 1) is updated accordingly.

<Modification>
The present invention is not limited to the embodiments described above, and various modifications are possible. The above-described embodiments are exemplified for easy understanding of the present invention, and are not necessarily limited to those having all the described configurations. Further, other configurations may be added to the configurations of the above embodiments, and part of the configurations may be replaced with other configurations. Also, the control lines and information lines shown in the drawings are those considered to be necessary for explanation, and do not necessarily show all the control lines and information lines necessary on the product. In practice, it may be considered that almost all configurations are interconnected. Possible modifications to the above embodiment are, for example, the following.

(1) In the above embodiment, an example in which a wind turbine generator is applied as the diagnosis target device 10 has been described. However, the diagnosis target device 10 is not limited to the wind turbine generator, and various devices such as industrial machines, electric vehicles, railroad vehicles, ships, elevators, and escalators can be applied as the diagnosis target device 10.

(2) In the above embodiment, a decision tree learning algorithm is used to calculate the importance levels Q1 to QN corresponding to the N measurement items. However, machine learning algorithms other than decision trees may be applied to calculate the importance levels Q1 to QN. For example, the teacher data DT may be learned by an algorithm such as random forest or support vector machine to obtain the degrees of importance Q1 to QN.

(3) In the above embodiment, the failure data DF (see FIG. 3) obtained from the diagnosis target device 10 itself is used to perform the abnormality diagnosis of the diagnosis target device 10 . However, for example, immediately after the diagnosis target device 10 is installed, there may be cases where there is no failure data DF or the amount of failure data DF is too small. Therefore, as the failure data DF, the failure data DF of another device having the same or similar specifications as the diagnosis target device 10 may be applied, and based on this, the teacher data DT and the like may be generated.

In particular, when the diagnosis target device 10 is a wind turbine generator, the “other device” from which the failure data DF is diverted is preferably another machine of the same wind farm. This is because the diagnosis target device 10 and the “other devices” have similar natural conditions such as wind speed, wind direction, and temperature. Further, the failure data DF, the teacher data DT, etc. may be corrected based on the difference in characteristics between the “other device” from which the failure data DF is diverted and the diagnosis target device 10 .

(4) In addition, in the above embodiment, one database 110 is implemented by one abnormality diagnosis device 100 . However, a plurality of abnormality diagnosis apparatuses 100 may be connected to a network (not shown) and the database 110 may be realized in storage on the network. Also, the functions of the abnormality diagnosis device 100 may be realized by performing distributed processing by a plurality of computers on a network.

(5) Since the hardware of the abnormality diagnosis device 100 in the above embodiment can be realized by a general computer, the flowchart shown in FIG. may be distributed through roads.

(6) The processing shown in FIG. 7 and the other processing described above have been described as software processing using a program in the above embodiment, but some or all of them are implemented in an ASIC (Application Specific Integrated Circuit). It may be replaced with hardware processing using an FPGA (Field Programmable Gate Array) or the like.

<Effect of the first embodiment>
As described above, the abnormality diagnosis apparatus 100 of the present embodiment includes the failure mode selection unit 106 that selects the failure mode DFM to be detected, and the failure mode selection unit 106 that is data corresponding to the failure mode DFM in the diagnosis target device 10 or another device. A teacher data creation unit 103 that creates teacher data DT for determining the presence or absence of a failure in the diagnosis target device 10 based on the corresponding measurement data DFK, and measurement items P1 to PN in the failure mode DFM based on the teacher data DT. A part of the measurement items P1 to PN is selected as a feature quantity (DQS) for the failure mode DFM based on the measurement item importance calculation unit 104 that calculates the importance Q1 to QN of and the calculated importance Q1 to QN. a feature amount selection unit 105, an anomaly degree calculation unit 107 for calculating an anomaly degree A corresponding to the failure mode DFM based on the measurement data DM related to the feature amount (DQS), and an anomaly degree calculation unit 107 based on the calculated anomaly degree A , and a device state determination unit 108 that determines the state of the diagnosis target device 10 .

According to the present embodiment, some of the measurement items P1 to PN are selected as feature quantities (DQS) for the failure mode DFM based on the calculated degrees of importance Q1 to QN. can be determined.

Further, it is preferable that the failure response measurement data DFK is data obtained from the diagnosis target device 10 .
This is because the failure response measurement data DFK from the device to be diagnosed 10 is considered to be highly compatible with the state of the device to be diagnosed 10 .

Further, the training data DT includes abnormal training data DT1 that presumes that an abnormality has occurred and normal training data DT2 that presumes normality. The failure response measurement data DFK from the discovery date and time DFT to the past by the first predetermined time (T1) is selected as the abnormal teaching data DT1, and immediately before the oldest (ts_c) data in the abnormal teaching data DT1 ( ts_b) to the second predetermined time (T2) in the past is selected as the normal teaching data DT2. is preferably a time set according to the failure mode DFM.

By setting the first and second predetermined times (T1, T2) according to the failure mode DFM in this way, the first and second predetermined times ( T1, T2) can be set.

10 diagnosis target device 100 abnormality diagnosis device (computer)
103 Teacher data creation unit (teacher data creation means)
104 measurement item importance calculation unit (measurement item importance calculation means)
105 feature quantity selection unit (feature quantity selection means)
106 failure mode selection unit (failure mode selection means)
107 Abnormality degree calculation unit (abnormality degree calculation means)
108 device state determination unit (device state determination means)
110 database (database means)
A Degree of anomaly DM Measured data DT Teacher data T1 Predetermined time (first predetermined time)
T2 predetermined time (second predetermined time)
DFK Failure response measurement data DFM Failure mode DFT Failure discovery date and time DQS Selected feature quantity (feature quantity)
DT1 Abnormal training data DT2 Normal training data P1 to PN Measurement items Q1 to QN Importance

Claims (4)

a database that stores measurement data of a plurality of measurement items in the device to be diagnosed;
a failure mode selection unit that selects a failure mode to be detected;
a teacher data creation unit that creates teacher data for determining the presence or absence of a failure in the device to be diagnosed based on failure response measurement data that is data corresponding to the failure mode in the device to be diagnosed or another device;
a measurement item importance calculation unit that calculates the importance of the measurement item in the failure mode based on the training data;
a feature amount selection unit that selects a part of the measurement items as a feature amount for the failure mode based on the calculated importance;
an anomaly degree calculation unit that calculates an anomaly degree corresponding to the failure mode based on the measurement data related to the feature amount;
and a device state determination unit that determines the state of the device to be diagnosed based on the calculated degree of abnormality. 2. The abnormality diagnosis device according to claim 1, wherein the failure response measurement data is data acquired from the diagnosis target device. The training data includes abnormal training data presumed to be abnormal and normal training data presumed to be normal,
The training data creation unit selects the fault response measurement data from the date and time of failure detection to the past by a first predetermined time as the fault training data, and selects the fault training data immediately before the oldest data among the fault training data. is selected as the normal teaching data, the failure response measurement data up to a second predetermined time in the past from the data of
3. The abnormality diagnosis device according to claim 2, wherein said first and second predetermined times are times set according to said failure mode. the computer,
database means for storing measurement data of a plurality of measurement items in the device to be diagnosed;
failure mode selection means for selecting a failure mode to be detected;
a teaching data creating means for creating teaching data for judging whether or not there is a failure in the device to be diagnosed based on measurement data corresponding to the failure mode in the device to be diagnosed or another device;
measurement item importance calculation means for calculating the importance of the measurement item in the failure mode based on the training data;
feature quantity selection means for selecting a part of the measurement items as a feature quantity for the failure mode based on the calculated degree of importance;
Anomaly degree calculation means for calculating an anomaly degree corresponding to the failure mode based on the measurement data related to the feature amount;
device state determination means for determining the state of the device to be diagnosed based on the calculated degree of abnormality;
A program to function as

Images