JP7211861B2 - Condition monitoring system and condition monitoring method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a condition monitoring system and a condition monitoring method for monitoring the condition of mechanical elements that constitute an apparatus.

2. Description of the Related Art Conventionally, in rotating machines, facilities, and plants including them, the states are monitored by measuring physical quantities using various sensors. For example, in a wind power generator, power is generated by rotating a main shaft connected to blades that receive wind power, increasing the speed of rotation of the main shaft with a gearbox, and then rotating a rotor of a generator. , a condition monitoring system (CMS) for diagnosing an abnormality in each part of such a wind power generator is known. In Japanese Patent No. 5917956 (Patent Document 1), diagnostic parameters measured by sensors that detect vibration, sound, etc., fixed to each part of a wind power generator are used to determine whether damage has occurred in each part. A condition monitoring system is disclosed for diagnosing whether

Japanese Patent No. 5917956

However, many types of anomalies can occur in such devices, such as damage to a wide variety of components, misalignment, imbalance about the axis of rotation, misalignment, and the like. In order to monitor all of these various abnormalities, it is necessary to perform processing by combining various types of diagnostic parameters with various signal processing (for example, filtering), which may require a lot of labor. .

SUMMARY OF THE INVENTION The present invention has been made to solve such problems, and its object is to provide a condition monitoring system and a condition monitoring method that can easily monitor various kinds of abnormalities.

According to one aspect of the present invention, a condition monitoring system includes a sensor and a processor, and monitors the condition of machine elements that make up the device. Sensors are used to detect the condition of mechanical elements. A processor is used to diagnose anomalies in mechanical elements. The processing device calculates a plurality of feature quantities from the detection signal of the sensor. The processing device extracts at least one feature amount whose temporal trends are independent of each other as an effective feature amount from the plurality of calculated feature amounts. The processing device diagnoses an abnormality in the machine element based on the extracted effective feature amount.

According to the present invention, it is possible to provide a condition monitoring system and a condition monitoring method that can easily monitor various kinds of abnormalities.

BRIEF DESCRIPTION OF THE DRAWINGS It is the figure which showed roughly the whole structure of the state-monitoring system which concerns on this Embodiment. 1 is a diagram schematically showing the configuration of a wind turbine generator; FIG. 4 is a table for explaining sensors and feature amounts regarding damage to each part of the wind turbine generator; 1 is a block diagram showing an overview of configurations of a monitor device and a control device according to an embodiment; FIG. It is a figure which explains notionally the function of the condition-monitoring system which concerns on a comparative example. 1 is a diagram conceptually explaining functions of a condition monitoring system according to an embodiment; FIG. It is a block diagram explaining the structure of the control apparatus regarding learning mode. 4 is a flowchart for explaining a method of calculating a change detection threshold; 4 is a flowchart for explaining a method of extracting effective feature amounts; It is a figure explaining the definition of an inflection point. FIG. 10 is a diagram illustrating an example of a method of calculating a time difference between change points of two feature amounts; FIG. 4 is a diagram for conceptually explaining a method of calculating a relevance evaluation value based on multiplication of second-order derivatives of a trend; FIG. 4 is a diagram for conceptually explaining feature amount normalization; It is a figure explaining an example of the Euclidean distance between feature-values. FIG. 10 is a flowchart for explaining a grouping method for integrating feature amounts in descending order of relevance; FIG. It is a figure explaining the concept of the grouping method which excludes from a feature-value with low relevance. FIG. 10 is a flowchart for explaining a grouping method for excluding features with low relevance. FIG. FIG. 18 is a diagram conceptually explaining the processing of steps S32 and S33 of FIG. 17; FIG. 10 is a diagram conceptually explaining duplicate evaluation values; It is a block diagram explaining the structure of the control apparatus regarding an operation mode. 10 is a flowchart for explaining processing of a re-extraction unit;

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The same or corresponding parts in the drawings are denoted by the same reference numerals, and the description thereof will not be repeated.

<Overall Configuration of Condition Monitoring System>
FIG. 1 is a diagram schematically showing the overall configuration of a condition monitoring system 100 according to this embodiment. Referring to FIG. 1 , condition monitoring system 100 monitors the condition of mechanical elements that make up wind turbine generator 10 . The status monitoring system 100 includes a monitor device 80 , a data server (monitoring side control device, hereinafter also simply referred to as a control device) 330 , and a monitoring terminal 340 . In the present embodiment, the wind turbine generator 10 or each part thereof is mentioned as a monitoring target of the state monitoring system according to the present embodiment, but this is only an example, and the monitoring target of the state monitoring system according to the present embodiment. is not limited to this.

The monitor device 80 is wired or wirelessly connected to sensors 70A to 70H (see FIG. 2), which will be described later. Based on the detection signals from the sensors 70A to 70H, the monitor device 80 calculates a "feature amount" indicating the characteristics of the detection signals. The feature quantity is, for example, an effective value, a peak value, a crest factor, an effective value after envelope processing, a peak value after envelope processing, or the like. The monitor device 80 transmits the feature quantity to the control device 330 via the Internet 320 .

Communication between the monitor device 80 and the control device 330 is performed by wire or wirelessly, and may be performed via the Internet as shown in FIG.

Control device 330 performs further detailed analysis based on the feature quantity received from monitor device 80 . In addition, the control device 330 determines whether or not each part of the wind turbine generator 10 is abnormal by statistical calculation of the feature amount. The control device 330 transmits the above results to the monitoring terminal 340 by wire or wirelessly. Control device 330 and monitoring terminal 340 are connected, for example, via an in-house LAN (Local Area Network) and/or the Internet. The control device 330 outputs feedback control, such as outputting an instruction of the feature amount to be observed to the monitor device 80, by wire or wirelessly as necessary (not shown).

The monitoring terminal 340 is used for viewing information related to the feature quantity received by the control device 330, instructing detailed analysis of the feature quantity, changing the settings of the monitor device 80, viewing the status of each device of the wind turbine generator, and the like. .

<Configuration of wind power generator>
FIG. 2 is a diagram schematically showing the configuration of the wind turbine generator 10. As shown in FIG. Referring to FIG. 2, wind turbine generator 10 includes main shaft 20, blades 30, gearbox 40, generator 50, main bearing 60, sensors 70A to 70H, and monitor device . Gearbox 40 , generator 50 , main bearing 60 , sensors 70A-70H and monitor device 80 are housed in nacelle 90 , which is supported by tower 91 .

Main shaft 20 enters nacelle 90 , is connected to an input shaft of gearbox 40 , and is rotatably supported by main bearing 60 . The main shaft 20 transmits the rotational torque generated by the blades 30 receiving the wind force to the input shaft of the gearbox 40 .

The blade 30 is provided at the tip of the main shaft 20 to convert wind power into rotational torque and transmit it to the main shaft 20 .

The main bearing 60 is fixed inside the nacelle 90 and supports the main shaft 20 rotatably. Main bearing 60 is configured by, for example, a rolling bearing such as a self-aligning roller bearing, a tapered roller bearing, a cylindrical roller bearing, or a ball bearing. These bearings may be single-row or double-row.

The sensors 70A to 70H are fixed to each piece of equipment inside the nacelle 90 and detect the states of the mechanical elements that make up the wind turbine generator 10 . Specifically, the sensor 70A is fixed on the upper surface of the main bearing 60 and detects the state of the main bearing 60 . Sensors 70B to 70D are fixed on the upper surface of speed increaser 40 and detect the state of speed increaser 40 . Sensors 70E and 70F are fixed on the upper surface of generator 50 and detect the state of generator 50 . Sensor 70G is fixed to main bearing 60 and detects misalignment and abnormal nacelle vibration. A sensor 70H is fixed to the main bearing 60 and detects unbalance and abnormal vibration of the nacelle. Hereinafter, the sensors 70A to 70H are also collectively referred to as "sensors 70". The sensor 70 detects vibration, sound, AE (Acoustic Emission), and the like for obtaining information of a monitoring target of the control device 330 . Alternatively, instead of the sensor 70, the value of the sensor 70 may be provided from a control device that collects the value of the sensor 70, such as SCADA (Supervisory Control And Data Acquisition).

The speed increaser 40 is provided between the main shaft 20 and the generator 50 , increases the rotation speed of the main shaft 20 and outputs it to the generator 50 . As an example, the speed increasing device 40 is configured by a gear speed increasing mechanism including a planetary gear, an intermediate shaft, a high speed shaft, and the like. It should be noted that, although not shown, a plurality of bearings for rotatably supporting a plurality of shafts are also provided in the gearbox 40 .

The generator 50 is connected to the output shaft of the speed increaser 40 and generates power by the rotational torque received from the speed increaser 40 . Generator 50 is configured by, for example, an induction generator. Bearings for rotatably supporting the rotor are also provided in the generator 50 .

The monitor device 80 is provided inside the nacelle 90 and receives detection signals such as vibration, sound, and AE of each device detected by the sensors 70A to 70H. Although not shown, the sensors 70A to 70H and the monitor device 80 are connected by wired cables or wirelessly.

<Relationship between diagnostic parameters and failure types>
FIG. 3 is a table for explaining sensors and feature amounts regarding damage to each part of the wind turbine generator. The control device according to the present embodiment compares the feature amount calculated from the detection signals of various sensors with a predetermined threshold to determine abnormality. Hereinafter, a threshold for judging abnormality when such a feature amount exceeds the value is also referred to as an "abnormality judgment threshold".

Specifically, for the main bearing 60 , the effective value is calculated by the monitor device 80 from the data measured by the high-frequency vibration sensor 70 A fixed to the main bearing 60 . If the controller 330 determines that the calculated effective value is equal to or greater than the corresponding abnormality determination threshold value, the monitoring terminal 340 displays that the main bearing 60 is damaged.

With respect to the main bearing 60, the data measured by the low-frequency vibration sensor 70H attached to measure the radial vibration of the main shaft is detected by the monitor device 80 as the primary rotation frequency component, secondary rotation frequency component, A tertiary rotation frequency component is calculated. When the controller 330 determines that each frequency component is equal to or greater than the corresponding abnormality determination threshold value, the monitoring terminal 340 displays that the main bearing 60 is unbalanced.

Furthermore, regarding the main bearing 60, the data measured by the low-frequency vibration sensor 70G attached to measure the axial vibration of the main shaft is detected by the monitor device 80 as the primary rotation frequency component, secondary rotation frequency component, A tertiary frequency component is calculated. When the controller 330 determines that each frequency component is equal to or greater than the corresponding abnormality determination threshold value, the monitoring terminal 340 displays that the main bearing 60 is misaligned with the main shaft 20 .

Further, regarding the gearbox 40, the effective value is calculated by the monitor device 80 from the data measured by the high-frequency vibration sensors 70B to 70D. If the controller 330 determines that the calculated effective value is equal to or greater than the corresponding abnormality determination threshold value, the monitoring terminal 340 displays that the bearing of the gearbox 40 is damaged.

Furthermore, regarding the gearbox 40, the primary meshing frequency component, the secondary meshing frequency component, and the tertiary meshing frequency component of the gear are calculated by the monitor device 80 from the data measured by the high-frequency vibration sensors 70B to 70D. . When the control device 330 determines that each calculated value is equal to or greater than the corresponding abnormality determination threshold value, the monitor terminal 340 displays that the gear of the gearbox 40 is damaged.

Furthermore, for the generator 50, the effective value is calculated by the monitor device 80 from the data measured by the high-frequency vibration sensors 70E and 70F. If the controller 330 determines that the calculated effective value is equal to or greater than the corresponding abnormality determination threshold value, the monitoring terminal 340 displays that the bearing of the generator 50 is damaged.

Furthermore, regarding the nacelle 90, the low-frequency vibration component is calculated by the monitor device 80 from the data measured by the low-frequency vibration sensor 70H attached to measure the radial vibration of the main shaft. When the control device 330 determines that the low-frequency vibration component is equal to or greater than the corresponding abnormality determination threshold value, the monitor terminal 340 displays that the nacelle 90 is vibrating abnormally.

As for the nacelle 90, the low-frequency vibration component is calculated by the monitor device 80 from the data measured by the low-frequency vibration sensor 70G attached so as to measure the vibration in the axial direction of the main shaft. When the control device 330 determines that the low-frequency vibration component is equal to or greater than the corresponding abnormality determination threshold value, the monitor terminal 340 displays that the nacelle 90 is vibrating abnormally.

The above measurement items have been partially extracted and explained for easy understanding, and are not limited to these. For example, for the measurement data of the vibration sensor, AE sensor, temperature sensor, and sound sensor, using a statistical method, the effective value, peak value, average value, crest factor, effective value after envelope processing, peak value after envelope processing may be calculated and compared to corresponding thresholds. With such a configuration, the control device 330 can monitor the state of each part of the wind turbine generator 10 , and the monitoring terminal 340 can display the state of each part.

FIG. 4 is a block diagram showing configurations of monitor device 80 and control device 330 according to the present embodiment. Referring to FIG. 4, monitor device 80 receives a signal from sensor 70 installed in wind turbine generator 10 . The control device 330 receives a signal from the monitor device 80, monitors the state of the wind turbine generator 10, and detects an abnormality.

Monitor device 80 includes A/D converter 110B, data acquisition section 120B, storage section 130B, data calculation section 140B, and output section 150B.

A/D converter 110B receives the detection signal of sensor 70 . The A/D converter 110B converts the detection signal, which is an analog signal, into a digital signal. The A/D converter 110B transmits the digital signal to the data acquisition section 120B.

The data acquisition unit 120B receives the digital signal from the A/D converter 110B, performs filtering, stores it in the storage unit 130B, and transmits it to the data calculation unit 140B.

The data calculation unit 140B receives the filtered detection signal of the sensor 70 from the data acquisition unit 120B and/or the storage unit 130B. The data calculation unit 140B calculates the feature amount from the detection signal of the sensor 70, stores it in the storage unit 130B, and transmits it to the output unit 150B.

The output unit 150B receives the feature amount from the data calculation unit 140B and transmits it to the data acquisition unit 120 of the control device 330 .

Control device 330 includes data acquisition section 120 , storage section 130 , data calculation section 140 , and output section 150 .

The data acquisition unit 120 receives the feature amount from the output unit 150 of the monitor device 80 , stores it in the storage unit 130 , and transmits it to the data calculation unit 140 .

The data calculation unit 140 reads out the feature amount calculated by the data calculation unit 140B from the storage unit 130 based on the detection signal of the sensor 70 when the wind turbine generator 10 is normal, and uses the read feature amount to determine if there is an abnormality using a statistical method. Generate a decision threshold. The abnormality determination threshold corresponds to an example of "first threshold". In this way, a mode for generating an abnormality determination threshold value for determining whether or not there is an abnormality using a feature amount is hereinafter referred to as a "learning mode" (details will be described later). Learning mode corresponds to one embodiment of "first mode."

Further, the data calculation unit 140 reads the detection signal of the sensor 70 during operation of the wind turbine generator 10 from the storage unit 130 and compares it with the abnormality determination threshold to determine whether the wind turbine generator 10 is abnormal or not. . The data calculation section 140 outputs the determination result to the output section 150 . A mode in which the state of the monitored object is monitored using the abnormality determination threshold is hereinafter referred to as an "operation mode" (details will be described later). The operational mode corresponds to one embodiment of the "second mode."

The output unit 150 receives the output of the data calculation unit 140 and transmits it to the monitoring terminal 340 .
<Definition of learning mode and operation mode>
Such a state monitoring system has a "learning mode" for generating a threshold value for determining whether or not a feature value is abnormal when the device to be monitored is normal; It has two modes, an "operational mode" that monitors the state of the device during operation using the thresholds generated in the learning mode.

That is, in the specification of the present application, the learning mode is a mode for generating an abnormality determination threshold value for determining whether or not an object to be monitored is abnormal using the feature amount when the wind turbine generator 10 is normal. , the operation mode is a mode in which the state of the monitored object is monitored using an abnormality determination threshold value.

The learning mode is typically performed immediately after a new device to be monitored by monitoring system 100 is installed. Immediately after the apparatus is newly installed, it can be assumed that each part of the apparatus is normal (no abnormality) because it is considered that each part of the apparatus is not out of order. In addition, when the status of the device to be monitored changes (for example, after confirming normal operation after replacing some parts of the device, or after performing regular maintenance and inspection), It may be configured to perform the learning mode as necessary. When the learning mode ends, the operating mode begins.

The following is a conceptual description of how controller 330 implements each mode. First, the configuration of the control device according to the comparative example will be described.

FIG. 5 is a diagram conceptually explaining functions of a condition monitoring system according to a comparative example. FIG. 5 includes a data acquisition unit 120 , a learning unit 1 and a monitoring unit 2 . Learning unit 1 and monitoring unit 2 are included in data computing unit 140 .

In the learning mode, the learning unit 1 receives feature amounts from the data acquisition unit 120 and uses the feature amounts to calculate an abnormality determination threshold value.

In the operation mode, the monitoring unit 2 receives feature amounts from the data acquisition unit 120 . The monitoring unit 2 compares the feature quantity with the abnormality determination threshold value calculated by the learning unit 1 . When the feature amount is equal to or greater than the abnormality determination threshold, the monitoring unit 2 determines that the feature amount is abnormal and outputs a warning.

However, in such a state monitoring system, when monitoring the entire device to be monitored, an abnormality determination threshold value is calculated for each component of the device and for each feature value that is an index of the type of abnormality of the component. However, it may be necessary to monitor whether each feature amount is equal to or greater than the abnormality determination threshold. However, in many cases, there are a wide variety of equipment components and their abnormalities monitored by such a condition monitoring system.

For example, many types of anomalies can occur in wind power plants, such as bearing and gear damage and mounting failures of many different shapes, unbalance and misalignment of the rotating shaft. Further, even if one abnormality is "abnormal number of rotations of the shaft", since the number of rotations of the shaft itself in the normal state is wide ranging from several tens of rpm to 1500 rpm, the number of rotations causing the abnormality also varies widely. In order to monitor such a wide variety of abnormalities without exception, it is possible to combine various calculation formulas and signal processing such as filtering to calculate a large number of feature values and abnormality determination thresholds (for example, 100 to 1000 each). may be required. Therefore, it may take a lot of effort for an analyst to confirm all of those feature quantities and abnormality determination thresholds.

<Operation of condition monitoring system according to the present embodiment>
FIG. 6 is a diagram conceptually explaining the functions of the condition monitoring system according to the present embodiment, and is a diagram to be compared with FIG. 6, the main difference between control device 330 according to the present embodiment and control device 330 (FIG. 5) according to the comparative example is that extraction unit 3 is added in learning mode, and selection unit 3 is added in operation mode. 5 and a change detection unit 4 are added. Extraction unit 3 , selection unit 5 and change detection unit 4 are included in data operation unit 140 .

In summary, state monitoring system 100 according to the present embodiment selects relevant A small number of effective feature quantities (for example, 10 to 50) are extracted by integrating feature quantities with high probability. Then, the state monitoring system 100 generates an abnormality determination threshold value based on these effective feature amounts. On the other hand, in the operation mode, the condition monitoring system 100 uses the abnormality determination threshold value to determine abnormality of effective feature amounts. This makes it possible to easily monitor various kinds of abnormalities.

Hereinafter, control device 330 according to the present embodiment will be described with an emphasis on functions different from those of control device 330 according to the comparative example, and description of the same functions will not be repeated.

<Overview of learning mode>
In the learning mode, the extraction unit 3 extracts a small number of effective feature amounts from a large number of feature amounts received from the data acquisition unit 120 and transmits the effective feature amounts to the learning unit 1 . In this specification, the term "effective feature amount" means a representative feature amount that can detect anomalies in a monitoring target by monitoring only those feature amounts among a large number of feature amounts, and/or a monitoring feature amount. It refers to the feature quantity to be monitored for target anomaly detection.

In order to determine whether each of the plurality of feature quantities received from the data acquisition unit 120 is "effective", the extraction unit 3 focuses on the presence or absence of change in each feature quantity. Specifically, the extraction unit 3 performs the following processes (1) to (3).

In the process (1), in order to determine whether or not there is a change in the feature amount, the control device 330 detects the change in the feature amount for all feature amounts based on the distribution of the feature amount during a certain period of time during the learning mode. is generated (hereinafter also referred to as "change detection threshold"). A change detection threshold corresponds to one embodiment of a "second threshold." The change detection threshold value may be manually set by the analyst based on information such as the feature amount read from the storage unit 130, or may be set by statistically processing the feature amount read from the storage unit 130. may be configured to be automatically calculated by

In the process (2), among all the feature quantities, the feature quantity whose value change is smaller than the change detection threshold value during the learning mode is determined as "a feature quantity that does not change (to be noticed)". and extracts the feature quantity as an “effective feature quantity”. A feature amount in which the change has not occurred corresponds to an example of the "first feature amount".

Note that when the value of the feature amount is smaller than the change detection threshold, for example, when the rotation speed of a rotating member is the feature amount, the change in the rotation speed over time (that is, the change in the value of the feature amount with respect to time, hereinafter referred to as (also called “trend”) shows a substantially constant value. The reason why a feature amount that has not changed is regarded as a valid feature amount will be explained after processing (3).

In the process (3), among all the feature amounts, the feature amount whose value is equal to or greater than the change detection threshold value during the learning mode is determined as "the feature amount in which a (notable) change has occurred". . The feature amounts that have undergone the change are integrated by evaluating correlation and/or similarity (details will be described later), and the integrated feature amount is extracted as an effective feature amount. The feature quantity in which the change has occurred corresponds to an example of the "second feature quantity".

Also, when the effective feature amount is extracted by the process (3), the extraction method for each effective feature amount is also stored. For example, assume that the first to fourth feature quantities F1(t) to F4(t) are integrated into one effective feature quantity, the fifth feature quantity F5(t), by the process (3). think. Assume that the fifth feature amount F5(t) in that case is represented by the function shown in the following formula (1) using the first to fourth feature amounts F1(t) to F4(t). where P(x) is a function of x.

F5(t)=P(F1(t), F2(t), F3(t), F4(t)) (1)
In this case, the extraction unit 3 causes the storage unit 130 to store the function represented by Equation (1). The function shown in formula (1) is used in the selection unit 5, which will be described later.

Here, the reason for extracting feature amounts that have not changed in process (2) as effective feature amounts will be described. It can be predicted that similar changes in timing and manner will occur in the future (in the operation mode) as well for the feature values that have changed in the learning mode. However, since it is not possible to predict how the feature values that have not changed in the learning mode will change in the future, it is necessary to continue to individually monitor the feature values that have not changed in the operation mode.

Furthermore, if the feature amount that has not changed is not used as a particularly effective feature amount in the process (2), the correlation and/or similarity is evaluated with the changed feature amount described in the process (3). This is because, if integration is performed, all the feature amounts in which the change has not occurred are combined into one feature amount in which the change has not occurred. However, as described above, it is not possible to predict the timing and mode of change in the event that a change occurs in the feature quantities that have not changed, so it is necessary to continue monitoring them individually.

Therefore, for the feature amounts that have not changed, before integrating the feature amounts in the process (3), each of them is treated as an effective feature amount in the process (2), so that individual monitoring is continued in the operation mode. and

On the other hand, it is assumed that the feature values that have changed are changed in the operation mode in the same way as in the learning mode. Therefore, it is conceivable that feature quantities that have undergone similar changes in the learning mode also undergo similar changes in the operation mode. For example, two feature quantities that show the same waveform in the learning mode are highly likely to continue to show the same waveform in the operation mode until an abnormality related to the feature quantity occurs. Therefore, by processing (3), if a plurality of feature amounts with similar changes are integrated into one effective feature amount in the learning mode, in the operation mode, by monitoring only the effective feature amount, In theory, it is thought that changes in multiple feature values can be detected.

Here, the feature quantities with similar changes are typically based on signals measured from the same sensors on a plurality of members having the same type, size, and operating state, or sensors installed at positions close to each other. , shows a plurality of feature quantities filtered in adjacent frequency bands before feature quantity calculation. The feature quantity of the plurality of members is, for example, the effective value of vibration of gears of the same size rotated at the same rotational speed and under the same load. In operational mode, the rms value is expected to remain the same until the associated anomaly occurs. For example, when an abnormality occurs in a bearing, vibration in a specific frequency band increases. Since the vibration propagates, the vibration is measured by a plurality of sensors installed in the vicinity of the bearing. Therefore, among the effective values (feature values) calculated from the signals of the nearby vibration sensors, all those that have been filtered so as to include the relevant frequency band increase. Therefore, for example, when there are 20 gears, when there are 3 vibration sensors in the vicinity of the 20 bearings, instead of monitoring the trend of the 3 feature values, one integrated Since it is possible to monitor the trend of effective feature amounts, the efficiency of monitoring is improved.

Conversely, two features with dissimilar changes are typically signals measured from sensors installed on a plurality of members of different types, sizes, and operating conditions, or sensors installed at locations separated from each other. and a plurality of features filtered in different frequency bands before feature calculation. Since these two feature amounts show trend graphs with different tendencies, if they are integrated, the original two feature amounts cannot be detected correctly. Therefore, these two feature amounts are not integrated with each other, but feature amounts with similar changes are appropriately integrated. Therefore, these two feature amounts are individually monitored even in the operation mode.

As will be described later, in this specification, the degree of similarity between feature quantities is indicated by the magnitude of the relationship evaluation value.

That is, in the learning mode, a large number of feature values indicating the state of the device to be monitored are appropriately integrated and extracted as a small number of effective feature values. , can detect various anomalies in the monitored device. That is, it is possible to efficiently monitor the state of the monitored device.

In this way, it is possible to monitor the entire monitoring target by extracting a small number of "effective feature values" to be monitored by processes (1) to (3) and monitoring only the effective feature values. become.

<Outline of operation mode>
Referring to FIG. 6 again, in the operation mode, selection unit 5 receives feature amounts from data acquisition unit 120 and extracts effective feature amounts from the received feature amounts. The selection unit 5 refers to a list of effective feature quantities extracted by the processing performed by the extraction unit 3, and performs integration processing to extract effective feature quantities. For example, the selection unit 5 selects the first to fourth feature quantities F1(t) to F4(t) by the above process (3) as one effective feature quantity using the above formula (1). 5 is integrated into the feature quantity F5(t).

In the operation mode, the monitoring unit 2 reads effective feature amounts from the storage unit 130 . The monitoring unit 2 compares the effective feature quantity with the abnormality determination threshold value calculated by the learning unit 1 . When the effective feature amount is equal to or greater than the abnormality determination threshold, the monitoring unit 2 determines that the effective feature amount is abnormal, that is, diagnoses that the mechanical element of the wind turbine generator 10 has an abnormality, Print a warning.

In the operation mode, the change detection unit 4 detects changes in effective feature amounts separately from the monitoring unit 2 . The change detection unit 4 determines that a feature amount whose trend change is greater than or equal to the change detection threshold value is a feature amount that has undergone a change, and executes re-extraction and re-learning of the feature amount.

<Detailed explanation of learning mode>
The learning mode will be described in more detail below.

The learning mode calculates an abnormality determination threshold and a change detection threshold based on the detection signal of the sensor 70 when the device to be monitored is in a normal state (for example, immediately after installation of the device). It is a mode to extract the amount.

FIG. 7 is a block diagram illustrating the configuration of control device 330 regarding the learning mode. In control device 330, data acquisition section 120, storage section 130 and data calculation section 140 (extraction section 3 and learning section 1) are used in the learning mode.

In the learning mode, the data acquisition unit 120 stores the trend of the feature amount acquired in the normal state in the storage unit 130 and transmits the trend to the extraction unit 3 .

The extraction unit 3 extracts an effective feature amount from the trend of the feature amount received from the data acquisition unit 120 by the above processes (1) to (3), and calculates a change detection threshold for the effective feature amount. . The extraction unit 3 transmits a list of effective feature quantities to the learning unit 1 . Extraction unit 3 transmits a list of effective feature quantities and the change detection threshold to storage unit 130 . Note that the change detection threshold value stored in the storage unit 130 is transmitted to the change detection unit 4 and used to detect changes in feature amounts in the operation mode (see FIG. 6).

The learning unit 1 calculates an abnormality determination threshold using a statistical method or the like based on the trend of the effective feature amount read from the storage unit 130 .

That is, the operation of control device 330 in the learning mode consists of the following operations (1) and (2).

Operation (1): The extraction unit 3 extracts "effective feature amounts" from trends of all feature amounts.
Operation (2): The learning unit 1 calculates an abnormality determination threshold from the trend of effective feature amounts.

Each of the operations (1) and (2) will be described in detail below.
8 and 9 are flow charts for explaining the operation (1).

FIG. 8 is a flowchart for explaining a method of calculating the change detection threshold, which is executed by the extraction unit 3. FIG.

In step S<b>01 , the extracting unit 3 reads the trend of the feature amount from the storage unit 130 . In step S02, the extracting unit 3 calculates the standard deviation σ and the average value μ of the feature amount trend. In step S03, the extraction unit 3 sets μ±r*σ as a change detection threshold, where r is a predetermined coefficient. where r is a positive number, for example 3. In step S04, extraction unit 3 stores the change detection threshold in storage unit 130, and ends the process.

FIG. 9 is a flow chart for explaining a method of extracting effective feature quantities, which is executed by the extraction unit 3 after the flow chart of FIG. In step S<b>11 , the extraction unit 3 reads the feature amount trend and the change detection threshold from the storage unit 130 . In step S12, the extraction unit 3 determines whether or not there is a feature amount whose trend change is smaller than the change detection threshold.

If there is a feature amount whose trend change is smaller than the change detection threshold (YES in step S12), in step S17, the extraction unit 3 selects the feature amount whose trend change is smaller than the change detection threshold as an effective feature amount. and In subsequent step S<b>18 , the extraction unit 3 stores effective feature amounts in the storage unit 130 .

On the other hand, the extracting unit 3 also extracts effective feature amounts in steps S13 to S16 from the feature amounts whose trend change is greater than or equal to the change detected in step S12, and stores them in the storage unit 130. FIG. In other words, in steps S13 to S16, the feature amount whose trend change is greater than or equal to the change detection threshold value in step S12 is processed.

After step S18, or when there is no feature amount whose trend change is equal to or greater than the change detection threshold value in step S12 (NO in step S12), the extraction unit 3 advances the process to step S13. In step S13, the extraction unit 3 calculates a "relevance evaluation value" that is an index for evaluating the relevance of the trend change of the feature amount for the feature amount that is equal to or greater than the change detection threshold in step S12. A relevance evaluation value is a value that evaluates the relevance of trend changes in feature quantities, and indicates the degree of similarity between two feature quantities. A method of calculating the relevance evaluation value will be described later in detail.

Subsequently, in step S14, the extraction unit 3 groups feature amounts using the relationship evaluation values. By this grouping, feature amounts with high relevance are put together in the same group. Grouping is a strategy for creating one group by units of similar feature amounts and later integrating the feature amounts within the group. A specific grouping method will be described in detail later.

In step S15, the extraction unit 3 extracts one feature amount from each group and sets it as an "effective feature amount". As the effective feature amount, a representative value may be selected from within the group, or a new feature amount may be calculated from the feature amounts within the group. By extracting effective feature amounts, the number of feature amounts that must be monitored is reduced, so that the efficiency of monitoring can be improved.

Subsequently, in step S17, the extraction unit 3 stores effective feature amounts in the storage unit 130, and terminates the process.

As a result of these processes, the storage unit 130 stores, as effective feature amounts, feature amounts that have not changed and feature amounts that have been extracted from the feature amounts that have changed.

<Relevance evaluation value>
Next, the relevance evaluation value will be described in detail. As described above, the relevance evaluation value is a value for evaluating the relevance of trend changes in feature quantities, and indicates the degree of similarity between two feature quantities. As examples of methods for calculating the relevance evaluation value, examples of the first to third calculation methods are given below. Also, a value obtained by integrating the values calculated by the first to third methods by arithmetic operations (multiplication, addition, etc.) may be used as the relevance evaluation value.

(1) First Calculation Method The first calculation method of the relevance evaluation value is a method of calculating the relevance evaluation value based on the time difference between the trend change points. Specifically, first, the trend of the first feature amount (hereinafter also referred to as the first trend) and the trend of the second feature amount (hereinafter also referred to as the second trend) are respectively changed points. calculate. Next, based on the time difference between the change point of the first trend and the change point of the second trend, the relationship between the two feature amounts is evaluated (detailed later). Note that the first calculation method is defined such that the smaller the time difference between the trend change points, the higher the relevance evaluation value. This is because, as will be described later, a threshold value is set for the relevance evaluation value, and if the relevance evaluation value is equal to or greater than the threshold value, processing is performed to determine that the first and second feature quantities are similar. be.

(1-1) Method of searching for a trend changing point First, as a method of searching for a trend changing point, for example, a second derivative of the trend is performed, and a point where the second derivative is 0, that is, a method of searching for an inflection point. There is

FIG. 10 is a diagram explaining the definition of an inflection point. Referring to FIG. 10, in the trend graph, the horizontal axis (t) indicates time and the vertical axis F(t) indicates the feature quantity. In the graph of y=F(t), the first derivative y=F'(t) of F(t) indicates the slope of the tangent line of y=F(t). Therefore, the second derivative F″(t) indicates the change in slope of the tangent line.

Here, attention is focused on the point at which the sign of F″(t) changes. Since F″(t) is the point at which the slope of the tangent line changes in sign, the slope of the tangent line changes from an increase to a decrease, or It is the point where the trend changes from a decrease to an increase, that is, the point where the trend changes (inflection point). Therefore, an inflection point is one of the representative examples of a trend change point. Therefore, in the first calculation method of the relevance evaluation value of this specification, the change point of the trend refers to the inflection point of the trend.

Next, as a method of calculating the evaluation value based on the time difference of the change points calculated by the above method, for example, (1-2) a method of dividing the sum of the reciprocals of the time differences of the change points by the number of change points, (1-3) A method of multiplying the second derivative of the trend (details will be described later) can be used. These two methods are described below.

(1-2) Method of dividing the sum of the reciprocals of the time differences of the change points by the number of change points Calculated by Next, the extraction unit 3 selects the one with the larger number of change points as the first feature amount F1(t) and the one with the smaller number of change points as the second feature amount F2(t ).

FIG. 11 is a diagram illustrating an example of a method of calculating the time difference between the change points of two feature quantities. The horizontal axis of FIG. 11 indicates time t, and the vertical axis indicates the first feature amount F1(t) and the second feature amount F2(t) at time t. The change points (hereinafter also referred to as first change points) of the first feature amount F1(t) are indicated as P11, P12, P13, . The change points (hereinafter also referred to as second change points) of the second feature quantity F2(t) are indicated as P21, P22, P23, .

Here, for each of the first changing points P11, P12, P13, . to calculate As shown in FIG. 11, the second change point closest to the first change point P11 is P21, so the time difference td1 is expressed as td1=|t21−t11| . Similarly, since the second change point closest to the first change point P12 is P22, the time difference td2=|t22−t12| of P12. Since the second change point closest to the first change point P13 is also P22, the time difference td3=|t22−t13| of P13.

Furthermore, the smaller the time difference td, the higher the relevance evaluation value, so the reciprocal of the time difference tr=1/td is taken. Next, this reciprocal tr is calculated for all first change points. Here, it is assumed that there are i_all first change points in total, and further, if the sum of tr for all ((i_all)) first change points is ts, then ts is given by the following equation (2) .

As a result, the smaller the time difference between the first change point and the second change point, the higher the sum ts. The higher is the relevance evaluation value. In this way, in order to cancel the influence of the number of first change points on the relevance evaluation value, if the value obtained by dividing the sum ts by the number of first change points is defined as the relevance evaluation value R1, then R1 is It is shown by following Formula (3).

R1=ts/(number of first change points) (3)
In this way, the relevance evaluation value R1 is calculated as the time difference between the change points of the first and second feature quantities F1(t) and F2(t) becomes smaller, that is, the trend changes of both feature quantities are closer. The stronger the tendency to occur over time, the higher the value. Therefore, the relevance evaluation value R1 indicates a larger value as the degree of similarity between the first and second feature quantities F1(t) and F2(t) is higher, based on the index of time when the trend change occurs. become.

(1-3) Method of Multiplying Second Derivative of Trend Next, a method of multiplying the second derivative of the trend will be described.

FIG. 12 is a diagram for conceptually explaining the method of calculating the relevance evaluation value based on the multiplication of the second derivative of the trend. It is assumed that the first feature amount F1(t) and the second feature amount F2(t) are trends having change points at the same time. Regarding the feature quantities F1(t) and F2(t), F1″(t)=F2″(t)=0 at time t=t11=t21 when the first change points P11 and P21 are detected. . Further, F1″(t)=F2″(t)=0 at time t=t12=t22 when the second change points P12 and P22 are detected. However, at t11<t<t12, |F1''(t)|>0 and |F2''(t)|>0.

Therefore, if the absolute value of the multiplication of the second derivative of the trend of the two feature values is r2=|F1″(t)×F2″(t)|, r2=0 at t=t11=t21, When t11<t<t12, r2>0.

Therefore, for two feature amounts having the same point of change, r2=0 at time t, which is both the point of change, and r2>0 at other times t.

On the other hand, referring to FIG. 12 again, the third feature amount F3(t) is a feature amount whose change point (P31) is detected at time t31 different from the feature amounts F1(t) and F2(t). .

Therefore, the absolute value r2 of the multiplication of the second derivative of the trends of the first feature amount F1(t) and the third feature amount F3(t) is r2 = 0 and r2>0 for other times t.

Therefore, for two feature amounts having different points of change, r2=0 at time t when one of them is the point of change, and r2>0 at other times t.

That is, in the two trends, the value of r2 tends to be smaller when the change points occur at different times than when the change points occur at the same time.

Therefore, assuming that the value obtained by adding the values of r2 for all times t (0≦t≦t_all) is the relevance evaluation value R2, R2 is expressed by the following equation (4).

That is, R2 indicates a larger value as the degree of similarity between the first and second feature quantities F1(t) and F2(t) is higher based on the trend change index. In this way, the extraction unit 3 can evaluate the relevance between the feature amounts by comparing the times of the change points of the temporal changes of the plurality of feature amounts.

(2) Second Calculation Method The second calculation method of the relevance evaluation value is a method in which the degree of similarity between trend changes of two feature quantities is used as the relevance evaluation value. That is, the higher the degree of similarity is, the higher the relevance evaluation value is set. Specifically, the degree of similarity is represented by, for example, the reciprocal of the Euclidean distance between normalized feature quantities.

(2-1) Normalization of Feature Amounts To explain in more detail, first, all feature amounts are normalized as a preparation for obtaining the Euclidean distance. As a method of normalization, for example, there is a method of making the maximum and minimum values of all feature quantities the same. For example, by setting the maximum value of all feature values to 1 and the minimum value to 0, all feature values can be normalized.

FIG. 13 is a diagram for conceptually explaining the normalization of feature amounts. Referring to FIG. 13, two feature quantities F1(t) and F2(t) show patterns that oscillate together, but differ in amplitude and baseline height. The normalized feature amounts f1(t) and f2(t ). Differences in amplitude and baseline height are canceled in f1(t) and f2(t) and only the vibration patterns can be compared.

Another normalization method is to use the median and the variance. For example, for all feature amounts, the median value of all time values in a predetermined period of each feature amount is calculated. Next, the median value and the variance can be normalized by using the value obtained by subtracting the median value from the value of each feature value at all times and further dividing it by the median value.

(2-2) Calculation of the Euclidean distance For the two feature values normalized in this way, the Euclidean distance between the feature values at each time is taken, and the reciprocal of the value obtained by adding the Euclidean distances at all times is related. It can be used as a property evaluation value.

FIG. 14 is a diagram illustrating an example of the Euclidean distance between feature quantities. Referring to FIG. 14, in Example 1, for example, feature quantities F1(t) and F2(t) whose behaviors are substantially identical are normalized, and as a result f1(t) and A graph of f2(t) is illustrated. In this case, if the difference |f2(t)−f1(t)| , the Euclidean distance is nearly zero. Therefore, assuming that the sum of the Euclidean distances for all times t (0≦t≦t_all) is S12, S12 is expressed by the following equation (5). Note that S12 is also the area between the graphs of f1(t) and f2(t).

That is, if the feature values have substantially the same behavior, the Euclidean distance for all t is small, so the value of S12 is also small.

On the other hand, referring to FIG. 14 again, Example 2 illustrates graphs of f1(t) and f3(t), which are results of normalizing the feature amounts F1(t) and F3(t). The behaviors of f1(t) and f3(t) are very different. In this case, the Euclidean distance |f3(t)-f1(t)| between the two graphs has a large value at most times t.

Therefore, assuming that the total value of the Euclidean distances for all t (0≦t≦t_all) is S13, S13 is expressed by the following equation (6). Note that S13 is also the area between the graphs of f1(t) and f3(t).

Therefore, the Euclidean distance is large on average between the feature quantities whose behaviors are significantly different from each other, so the value of S13 is also large.

Therefore, the value of S decreases as the similarity of the trend change of the feature amount increases. Therefore, if the relevance evaluation value R3=1/S, R3 indicates a larger value as the similarity of trend changes is higher. In this way, the extraction unit 3 can evaluate the relevance by normalizing a plurality of feature quantities and comparing the reciprocals of the differences between the normalized plurality of feature quantities.

A third method of calculating the relevance evaluation value is a method of using the absolute value of the normalized correlation coefficients of the two feature quantities as the relevance evaluation value. Note that the normalized correlation coefficient is known as one of the indices indicating the strength of the relationship between two variables.

As preparation for calculating the relevance evaluation value, first, average values F1(t)ave, F2( t) ave is calculated by the following equations (7) and (8).

Here, assuming that the normalized correlation coefficient is I, I is expressed by the following equation (9).

Assuming that the absolute value of this normalized correlation coefficient I is the relevance evaluation value R3, R3=|I|.
As mentioned above, the normalized correlation coefficient is an indicator of the strength of the relationship between two variables. Therefore, R3 indicates a larger value as the similarity between trend changes of the two feature amounts increases. In this way, the extraction unit 3 can evaluate the relevance of feature amounts based on the normalized correlation of the feature amount trends.

As described above, by using any one of the first to third methods, the extraction unit 3 calculates the relationship evaluation value so that the evaluation value increases as the feature amounts are more similar to each other. can be done. Note that the definition and calculation method of the relationship evaluation value are not limited to the above methods, and any value for evaluating the relationship between trend changes between feature quantities and a method capable of calculating the value may be used.

<Grouping>
Next, grouping will be described in detail. In this specification, the purpose of grouping is to put feature quantities with high relevance evaluation values into the same group. There are, for example, the following two methods for such grouping.

A first grouping method is a method of integrating feature amounts in descending order of relevance. The processing flow is shown in FIG.

(1) Grouping Method for Integrating Feature Amounts with High Relevance First FIG.

By step S<b>20 , the extraction unit 3 reads out the relevance evaluation value regarding the feature amount from the storage unit 130 . In step S21, the extraction unit 3 selects a combination of feature quantities with the highest relevance evaluation value.

In step S22, the extracting unit 3 determines that the relevance evaluation value of the feature quantity with the highest relevance evaluation value (hereinafter also referred to as "maximum relevance evaluation value") is equal to or greater than the relevance threshold value. Determine whether or not If the maximum relevance evaluation value is greater than or equal to the relevance threshold value (YES in step S22), the extraction unit 3 integrates the feature amounts in step S24. As a method for the integration, a method of averaging the two feature amounts after normalizing the two feature amounts can be used. Other examples of the integration method include a method of discarding one of the two feature amounts as a representative value and discarding the other feature amount, It is possible to adopt a method of converting

After step S24, in step S25, the extraction unit 3 recalculates the relationship evaluation value between the integrated feature amount and other feature amounts. Subsequently, the extraction unit 3 returns to the process of step S20 again, and repeats steps S20 to S22 until the maximum value of the relevance evaluation values becomes smaller than the relevance threshold value.

When the maximum relevance evaluation value is smaller than the relevance threshold value, it means that all the remaining feature quantities have low relevance to each other, that is, they are independent feature quantities each having its own individuality. show. It also indicates that feature quantities having high relevance, that is, having similar individuality have already been integrated. Therefore, when the maximum relevance evaluation value is smaller than the relevance threshold value, all the remaining feature amounts are considered to be "effective feature amounts" that are desirably monitored in the operational mode. In addition, this relevance threshold is configured so that the analyst can set an appropriate value in consideration of the above purpose, or it is automatically calculated based on the feature amount by statistical processing by computer. preferably configured.

Therefore, in step S22, if the maximum relevance evaluation value is smaller than the relevance threshold value (NO in step S22), extraction unit 3 converts the feature quantity remaining in storage unit 130 to a valid feature quantity in step S24. , and terminate the process. In this way, the extracting unit 3 calculates effective feature amounts by grouping feature amounts with high relevance in order.

(2) Grouping Method for Removing Features with Low Relevance FIG. 16 is a diagram illustrating the concept of a grouping method for removing features with low relevance.

(2-1) Referring to FIG. 16, extraction unit 3 first groups all the feature amounts. Here, 10 feature amounts F1 to F10 are exemplified. In addition, the feature amounts with high similarities, that is, feature amounts with high relevance evaluation values are indicated by circles, triangles, and squares. That is, the feature quantity F1 indicated by the circle has a relatively high degree of similarity with the feature quantity F3 indicated by the same circle (for example, the relevance evaluation value is equal to or greater than the relevance threshold value), but is indicated by the triangle. The similarity with the feature F2 obtained is relatively low (for example, the relevance evaluation value is less than the relevance threshold).

(2-2) Next, the extraction unit 3 groups each related feature quantity. As will be described in detail with reference to FIG. 18, here, the process of duplicating the group and removing the two least relevant feature quantities from the duplicated group is repeated. As a result, as shown in the figure, it is possible to generate a plurality of groups in which the relevance evaluation values of the two feature quantities with the lowest relevance within the group are higher than those of the original group. Here, the groups G1 to G5 are arranged in descending order of the number of feature values in the group.

However, in the result of (2-2), the same element exists in multiple groups (for example, in FIG. 16, both groups G1 and G5 have feature F5, and both groups G3 and G4 have feature F8). ), and so on.

(2-3) Therefore, the extraction unit 3 integrates groups having overlapping feature amounts. As a result, groups with a high duplication of feature amounts are integrated.

(2-4) Finally, the extracting unit 3 integrates the feature amounts in each group to obtain "effective feature amounts". For example, in FIG. 16, the effective feature amount Fa is calculated by integrating the feature amounts F1, F3, F5, and F6 of the group of circles. Since the features F1, F3, F5 and F6 are similar to each other, the effective feature Fa has the same characteristics as the features F1, F3, F5 and F6.

FIG. 17 is a flowchart for explaining specific processing of the grouping method for removing feature quantities with low relevance, which is executed by the extraction unit 3 . In step S<b>30 , the extraction unit 3 reads the relevance evaluation value, the relevance threshold value, and the duplication threshold value from the storage unit 130 .

(2-1) Group All Feature Amounts In step S31, the extraction unit 3 groups all feature amounts into one group.

(2-2) Grouping by Related Feature Amounts In steps S32 and S33, the extraction unit 3 groups by related feature amounts. For this purpose, the extraction unit 3 performs a process of duplicating the group and increasing the relevance evaluation value within the group.

In step S32, the extraction unit 3 duplicates the group. In step S<b>33 , the extraction unit 3 deletes the feature quantities of the combination with the lowest relevance evaluation value one by one in the two groups (detailed with FIG. 18 ).

FIG. 18 is a diagram conceptually explaining the processing of steps S32 and S33 of FIG. FIG. 17 illustrates the first processing of steps S32 and S33 and the second processing of steps S32 and S33 after returning from step S37.

The processing of steps S32 and S33 for the first time will be described with reference to FIG. In step S32, one group obtained by grouping all the feature quantities F1 to Fn generated in step S31 is duplicated. In this specification, hereinafter, n and m refer to natural numbers. In step S33, the two feature quantities with the lowest relevance evaluation values in the group before duplication are deleted one by one in the two groups after duplication. For example, in FIG. 18, when the two feature quantities with the lowest relevance evaluation values are F1 and Fn, the feature quantity F1 is deleted from one group and the feature quantity Fn is deleted from the other group. By processing in this way, the relevance evaluation value within the group increases while all the feature values are present in some group.

Referring to FIG. 17 again, in subsequent steps S34 to S37, extraction unit 3 isolates groups with high relevance evaluation values.

In step S34, the extraction unit 3 "isolates" a group having one feature amount in the group. In the grouping for each related feature amount, isolation indicates that the process is not performed up to step S38 and is used again in step S38. In step S35, the extraction unit 3 calculates the average value of the relevance evaluation values within the group. In step S36, the extraction unit 3 also isolates groups whose average relevance evaluation values are equal to or greater than the relevance threshold value. In step S37, the extracting unit 3 determines whether or not there is a non-isolated group. If there is a group that is not isolated (YES in step S37), the extraction unit 3 returns the process to step S32.

Therefore, as a result of steps S34 to S37, a group with a high relevance evaluation value of feature quantities in the group, that is, a group containing only feature quantities with a high degree of similarity is isolated. Therefore, the processing for increasing the similarity of the feature amounts in the groups in steps S32 and S33 is performed again for groups with low similarities in the feature amounts within the group.

The processing of steps S32 and S33 for the second time will be described using FIG. 18 again. In step S32, each of the groups remaining in step S37 is duplicated. In step S33, the two feature quantities with the lowest relevance evaluation values in the group before duplication are deleted one by one in the two groups after duplication. For example, if the groups generated in the second step S32 in FIG. The feature amount F2 is deleted from the group G2, and the feature amount Fn is deleted from the group G2. On the other hand, when the two feature quantities with the lowest relevance evaluation values in the group G3 are F1 and F(n-1), the feature quantity F1 from the group G3 and the feature quantity F(n-1) from the group G4 are erased respectively. That is, as a result of repeating steps S32 and S33 and removing the grouped feature quantities in descending order of relevance, groups of similar feature quantities are formed as shown in FIG. can.

(2-3) Integrate groups having overlapping feature amounts.
Next, in steps S38 to S48, the extraction unit 3 integrates groups having overlapping feature amounts.

When all the groups are isolated in step S37 (NO in step S37), in step S38, the extraction unit 3 classifies the groups as G1 to Gn in descending order of the number of elements (that is, the number of feature amounts included in the group). do. In step S39, the extraction unit 3 sets m=1.

In step S40, the extraction unit 3 determines whether or not the overlap evaluation value |Gn∩Gm|/|Gn| is equal to or greater than the overlap threshold. Here, |Gn∩Gm| indicates the number of feature amounts commonly included in groups Gn and Gm, and |Gn| indicates the number of feature amounts included in group Gn.

FIG. 19 is a diagram conceptually explaining the overlap evaluation value. In Example 1, let Gn={F1, F2, . . . , F10} and Gm={F1, F2, . That is, Gn∩Gm={F1, F2, . In this case, |Gn|=10 and |Gn∩Gm|=9, so the overlap evaluation value |Gn∩Gm|/|Gn| is 0.9. Note that the overlap threshold is a value that can be manually set in advance by the analyst, and is, for example, 0.9. Alternatively, the overlap threshold value may be automatically calculated from the overlap evaluation values of the remaining groups by performing statistical processing or the like. From the above, if |Gn∩Gm|/|Gn| is equal to or greater than the overlap threshold, it means that the degree of overlap of the elements included in group Gn is high.

If |Gn∩Gm|/|Gn| is equal to or greater than the duplication threshold (YES in S40), the extraction unit 3 integrates the group Gn into the group Gm in step S41. More precisely, elements (feature amounts) that do not overlap with groups Gm included in group Gn are added to group Gm.

Again referring to Example 1 of FIG. 19, since duplicate evaluation value |Gn∩Gm|/|Gn|=0.9≧duplicate evaluation value, group Gn is integrated into group Gm. Therefore, F10, which is a feature amount that does not overlap with groups Gm included in group Gn, is added to group Gm.

On the other hand, if |Gn∩Gm|/|Gn| is less than the overlap threshold (NO in S40), extraction unit 3 does not integrate groups Gn, and proceeds to step S42.

Example 2 of FIG. 19 shows an example of Gn={F1, F2, . . . , F10} and Gm={F1, F2, . That is, it differs from Example 1 in that the feature amount F10 is not included in |Gn∩Gm|. Therefore, |Gn∩Gm|/|Gn|=0.8, which is less than the overlap threshold. In such cases, groups Gn are not merged.

In step S42, the extraction unit 3 determines whether or not m is (n-1) or less. That is, it is checked whether the group Gn overlaps with each of the groups G1 to G(n-1), and if there is an overlap, it is determined whether processing such as integration has been performed.

If m is (n−1) or less (YES in S42), in step S43, extraction unit 3 adds 1 to the value of m, and returns the process to step S40.

If m is greater than (n-1) in step S42 (NO in S42), in step S44, the extraction unit 3 determines whether or not the group Gn has been integrated with another group even once. If the group Gn has been integrated with another group even once (YES in step S44), the extraction unit 3 deletes the group Gn in step S45, and proceeds to step S47. If group Gn has not been integrated with another group even once (NO in step S44), extraction unit 3 isolates group Gn in step S46, and then proceeds to step S47. In other words, the group Gn is an independent group that has many elements that do not overlap with other groups.

In step S47, the extraction unit 3 determines whether or not n is 1 or less. When n is greater than 1 (NO in step S47), extraction unit 3 reduces the value of n by 1 in step S48, and returns the process to step S38.

Therefore, in steps S38 to S48, groups with a high duplication degree are integrated, and only groups with independent characteristics are isolated and remain.

(2-4) Integrate the feature amounts in each group and set them as "effective feature amounts".
When n is 1 or less (YES in step S47), the number of groups is 1 or less, and therefore cannot be integrated. Therefore, the extraction unit 3 calculates an effective feature amount for each group in the last remaining group G1 and the group isolated in step S49. Effective feature amounts are integrated by means of addition, averaging, selection of representative values, or the like, as described above. In subsequent step S50, the extraction unit 3 stores the effective feature amount in the storage unit 130, and ends the process.

In step S49, a feature amount is calculated for each group having a feature amount with a low duplication degree. Note that the elements (feature amounts) within each group are only similar to each other in the processes up to step S37. Therefore, the number of groups from which the feature values each having individuality are isolated is calculated. In other words, the extraction unit 3 extracts at least one feature amount whose temporal trends are independent from each other as an effective feature amount from the plurality of feature amounts input to the extraction unit 3 .

In the processing loop of steps S40 to S43, the group Gn is integrated into all the groups Gm whose overlap evaluation value |Gn∩Gm|/|Gn| Groups having a common feature amount are integrated in such a way that if the threshold value is exceeded, the groups are integrated again. However, the method of integrating groups having common feature amounts is not limited to the above example, and any method that achieves the goal of appropriately integrating groups having common feature amounts may be used. For example, it is also possible to adopt another method such that the group Gn is integrated only into the group Gm in which |Gn∩Gm|/|Gn| is the largest.

As described above, in the learning mode of the state monitoring system 100 according to the present embodiment, a trend change is detected from a large amount of feature values detected by a large number of sensors that monitor the behavior of each part of the device to be monitored. The feature values above the threshold value are (1) grouped so that feature values with high trend relevance fall into the same group, and (2) the feature values are integrated within the group to obtain the trend of each group. A small number of valid features that characterize the change are extracted. Therefore, in the operation mode described below, it is possible to detect various abnormalities in the monitoring target device simply by monitoring the effective feature amount.

<Detailed description of operation mode>
Next, the operation modes will be explained in detail. In the learning mode, the controller 330 extracted a small number of valid features from a large number of features. In the operation mode, the control device 330 monitors the state of the monitored device by monitoring the small number of effective feature values. In operational mode, controller 330 analyzes the detection signals obtained from sensor 70 . FIG. 20 shows the configuration of the operation mode.

FIG. 20 is a block diagram illustrating the configuration of the control device 330 regarding the operation mode. In the control device 330, the data acquisition unit 120, the storage unit 130, the data calculation unit 140 (the selection unit 5, the change detection unit 4, the re-extraction unit 6, the re-learning unit 7, and the monitoring unit 2) and the output unit 150 operate in the operation mode. used.

The selection unit 5 reads out the feature amount and the effective feature amount list from the storage unit 130 . The selection unit 5 selects or integrates and extracts effective feature amounts from all the feature amounts based on the effective feature amount list.

The change detection unit 4 reads the effective feature quantity from the selection unit 5 and the change detection threshold value from the storage unit 130 . The change detection unit 4 determines, from among the effective feature amounts, an effective feature amount whose change over time is equal to or greater than the change detection threshold as a "feature amount in which a (remarkable) change has occurred". is output to the re-extraction unit 6.

The re-extraction unit 6 receives the list of changed effective feature amounts from the change detection unit 4 and reads the changed effective feature amount trend from the storage unit 130 . The re-extraction unit 6 performs re-extraction for extracting a new effective feature amount from the effective feature amount for which a change has been detected. The re-extraction unit 6 transmits the new effective feature amount list to the re-learning unit 7 and the storage unit 130 .

The re-learning unit 7 receives the list of effective feature amounts from the re-extraction unit 6 and reads out the trend of effective feature amounts from the storage unit 130 . The re-learning unit 7 performs the same processing as the learning unit that is performed in the learning mode for the new effective feature quantity, and calculates the abnormality determination threshold value. Re-learning unit 7 stores the abnormality determination threshold in storage unit 130 .

Storage unit 130 adds the abnormality determination threshold value to the abnormality determination threshold value calculated by learning unit 1 .

The monitoring unit 2 reads out the abnormality determination threshold value from the storage unit 130 and compares the effective feature amount with the abnormality determination threshold value. The monitoring unit 2 determines that a feature amount that is equal to or greater than the abnormality determination threshold value is abnormal. The monitoring unit 2 outputs the determination result to the output unit 150 .

When the output unit 150 receives the determination result that an abnormality has occurred, the monitoring terminal 340 warns the analyst and/or the monitoring system and/or the Notify the type of the abnormality.

FIG. 21 is a flowchart for explaining the processing of the re-extraction unit, which is executed by the re-extraction unit 6. FIG.

In step S<b>60 , the re-extraction unit 6 reads the feature amount trend and the change detection threshold from the storage unit 130 . This trend is a trend for any period of time, including the most recent trend. For example, use a trend accumulated in operational mode that is as long as the period of the trend used in learning mode.

In step S61, the re-extraction unit 6 collects feature amounts equal to or greater than the change detection threshold, and calculates a relevance evaluation value. The calculation method of the relevance evaluation value is the same as in the learning mode.

In step S62, the re-extraction unit 6 groups the feature amounts using the relevance evaluation value. The grouping method is the same as in learning mode.

In step S63, the re-extraction unit 6 extracts "effective feature amounts" from the feature amounts of each group. The method of extracting effective feature quantities is the same as in the learning mode. In step S63, the re-extraction unit 6 stores the effective feature amount in the storage unit 130, and terminates the process.

As described above, according to the control device according to the present embodiment, effective feature quantities can be extracted from a wide variety of feature quantities calculated by the monitor device 80 in the learning mode. As a result, it is possible to easily set and manage threshold values for detecting anomalies in feature quantities, and to monitor devices to be monitored using the threshold values. Further, according to the control device according to the present embodiment, it is possible to detect feature amounts that have changed during the operation mode, re-extract effective feature amounts, and re-calculate the abnormality determination threshold value. . As a result, even if the operation status of the monitored device changes during the operation mode, it is possible to prevent an abnormality from being overlooked. Therefore, it is possible to provide a condition monitoring system and a condition monitoring method that can easily monitor various kinds of abnormalities.

The embodiments disclosed this time should be considered as examples and not restrictive in all respects. The scope of the present invention is indicated by the scope of the claims rather than the description of the above-described embodiments, and is intended to include all modifications within the scope and meaning equivalent to the scope of the claims.

1 learning unit 2 monitoring unit 3 extraction unit 4 change detection unit 5 selection unit 6 re-extraction unit 7 re-learning unit 10 wind turbine generator 20 main shaft 30 blades 40 gearbox 50 power generator , 60 main bearing, 70A to 70F high frequency vibration sensor, 70G, 70H low frequency vibration sensor, 80 monitor device, 90 nacelle, 91 tower, 100 condition monitoring system, 110 A / D converter, 120, 120B data acquisition unit, 130, 130B storage section, 140, 140B data calculation section, 150, 150B output section, 320 Internet, 330 monitoring side control device (data server), 340 monitoring terminal.

Claims (12)

A condition monitoring system for monitoring the condition of mechanical elements that make up a device,
a sensor for detecting the state of the mechanical element;
a processing device for diagnosing an abnormality in the mechanical element;
The processing device is
calculating a plurality of feature quantities from the detection signal of the sensor;
Extracting an effective feature quantity from the plurality of calculated feature quantities,
diagnosing an abnormality in the mechanical element based on the extracted effective feature quantity;
extracting, from among the plurality of calculated feature amounts, a first feature amount whose change over time is smaller than a change detection threshold value as the effective feature amount;
Among the plurality of calculated feature quantities, a second feature quantity whose change over time is equal to or greater than the change detection threshold is evaluated for at least one of correlation and similarity, integrated, and integrated. A condition monitoring system for extracting a feature quantity as the effective feature quantity . said processor configured to sequentially perform first and second modes;
the first mode is performed when the device is normal;
wherein, in the first mode, the processing device extracts the effective feature amount based on the temporal tendency of the plurality of feature amounts calculated from the detection signal of the sensor;
2. The condition monitoring system according to claim 1, wherein said processing device diagnoses an abnormality of said mechanical element in said second mode based on said effective feature quantity during operation of said device. The processing device is
generating an abnormality determination threshold using a statistical method from the effective feature amount in the first mode;
3. The condition monitoring system according to claim 2, wherein in said second mode, when said valid feature amount is equal to or greater than said abnormality determination threshold value, said machine element is diagnosed as being abnormal. During the second mode, the processing device detects, from the effective feature amounts, a feature amount whose change over time is equal to or greater than the change detection threshold, and extracts the effective feature amount from the detected feature amount. 4. The condition monitoring system according to claim 3, wherein the quantity is newly generated, and the abnormality determination threshold value is also generated for the newly generated effective feature quantity. The processing device, in the first mode,
performing a grouping process of grouping the second feature amount based on the relevance of changes over time ;
5. The condition monitoring system according to any one of claims 2 to 4 , wherein a feature amount characterizing the change over time of each group is extracted as said effective feature amount. 6. The condition monitoring system according to claim 5, wherein said processing device evaluates said relevance based on a normalized correlation of changes over time between said plurality of second feature quantities. 6. The condition monitoring system according to claim 5, wherein said processing device evaluates said relevance by comparing times of change points of changes over time of each of said plurality of second feature quantities. The processing device evaluates the relevance by normalizing each of the plurality of second feature quantities and comparing reciprocals of differences between the plurality of normalized second feature quantities. 6. The condition monitoring system according to 5. The condition monitoring system according to any one of claims 5 to 8, wherein in the grouping process, the processing device groups the plurality of second feature quantities in descending order of relevance. 9. The state monitoring according to any one of claims 5 to 8, wherein in the grouping process, the processing device removes the grouped plurality of second feature quantities in descending order of relevance. system. 11. The processing device according to any one of claims 2 to 10, wherein said processing device sets said change detection threshold value based on a distribution state of said plurality of feature quantities during a certain period of time during said first mode. Condition monitoring system. A condition monitoring method for monitoring the condition of a machine element that constitutes an apparatus, comprising:
detecting the state of the machine element with a sensor installed in the device;
a step of calculating a plurality of feature quantities from the detection signal of the sensor;
a step of extracting an effective feature quantity from the plurality of calculated feature quantities;
diagnosing an abnormality of the mechanical element based on the extracted effective feature quantity ;
In the extracting step,
Among the plurality of calculated feature amounts, for a first feature amount whose change over time is smaller than a change detection threshold value, the first feature amount is extracted as the effective feature amount,
Among the plurality of calculated feature quantities, a second feature quantity whose change over time is equal to or greater than the change detection threshold is evaluated for at least one of correlation and similarity, integrated, and integrated. A condition monitoring method , wherein a feature quantity is extracted as the effective feature quantity .

Images