JP6791770B2 - Status monitoring method and status monitoring device - Google Patents

Description

The present invention relates to a condition monitoring method and a condition monitoring device.

Conventionally, rotating machines and equipment and plants including them monitor the state by measuring physical quantities using various sensors.

As a method of condition monitoring, there is a method of creating a model of a normal state of a physical quantity and determining an abnormality by calculating how much the newly measured physical quantity deviates from the model of the normal state (for example, Patent Document). See 1). There is also a method of storing the value of the characteristic frequency peak generated by bearing damage or shaft runout in a condition monitoring system and identifying an abnormality by observing the change of the peak (see, for example, Patent Document 2).

Japanese Patent No. 5431235 Japanese Patent No. 5780870

However, in a rotating machine that is affected by operating conditions and noise, the physical quantity fluctuates even in the normal state, so the physical quantity indicating an abnormality is buried in the model in the normal state, or it is normal depending on the created model in the normal state. There is a possibility that the condition will be mistaken for abnormal.

For rotating machines affected by operating conditions and noise, a condition monitoring method that can eliminate erroneous discrimination as much as possible and discriminate minute damage is desired.

The present invention has been made to solve the above problems, and an object of the present invention is to provide a condition monitoring method and a condition monitoring device having a reduced erroneous discrimination rate.

The present invention is a state monitoring method, wherein when the object to be tested is normal, a plurality of first measurement data having a length of one time are acquired from a sensor installed on the object to be tested at different timings. One step, a second step of acquiring a plurality of second measurement data having a length of the first time at different timings at the time of diagnosing the object to be tested, and a plurality of learning data randomly selected from the plurality of first measurement data. The third step of selecting the data, the fourth step of randomly selecting a plurality of test data from the plurality of second measurement data, and each of the plurality of training data for each second hour length shorter than the first time length. The fifth step of dividing into segment data and creating a first feature quantity vector containing a plurality of feature quantities for each segment data after division, and a plurality of first steps created for each segment data for a plurality of training data. From the feature quantity vector of 1, the sixth step of creating the classification boundary for classifying normal and abnormal and the abnormality discrimination threshold, and each of the plurality of test data are divided into segment data for each second time length. The seventh step of creating a second feature quantity vector containing a plurality of feature quantities for each segment data after division, and the anomaly degree which is the distance from the classification boundary for the second feature quantity vector are calculated. , Eighth step of calculating the abnormality rate in which the number of abnormalities of the second feature quantity vector exceeding the abnormality discrimination threshold is the ratio to the total number of the second feature quantity vectors for each of the plurality of test data. And a ninth step of repeatedly executing the third to eighth steps a plurality of times and determining the object to be tested as an abnormality when the average value of the obtained abnormality rates exceeds a predetermined value.

The state monitoring method in another aspect of the present invention acquires a plurality of first measurement data having a first time length from a sensor installed on the test object at different timings when the test object is normal. A plurality of learning data are obtained from the first step, the second step of acquiring a plurality of second measurement data having a length of the first time at different timings at the time of diagnosing the object to be tested, and the plurality of first measurement data. The third step of randomly selecting, the fourth step of randomly selecting a plurality of test data from a plurality of second measurement data, and each of the plurality of training data for each second hour length shorter than the first time length. A first feature quantity vector containing a plurality of feature quantities is created for each segment data after the division, and the first feature quantity vectors of a plurality of consecutive segment data are put together to form a first feature quantity vector. From the fifth step of creating the feature quantity vector of 3, the classification boundary for classifying normal and abnormal from the plurality of third feature quantity vectors created for each of a plurality of consecutive segment data for the plurality of training data, and The sixth step of creating the abnormality discrimination threshold, and the second step of dividing each of the plurality of test data into segment data for each second time length and including a plurality of feature amounts for each segment data after the division. The seventh step of creating a feature quantity vector and combining the second feature quantity vectors of a plurality of consecutive segment data to create a fourth feature quantity vector, and the fourth feature quantity vector from the classification boundary. The degree of anomaly, which is the distance, is calculated, and for each of the plurality of test data, the number of the degree of abnormality of the fourth feature amount vector exceeding the abnormality determination threshold value is the ratio to the total number of the fourth feature amount vectors. The eighth step of calculating the abnormality rate and the third to eighth steps are repeatedly executed a plurality of times, and when the average value of the obtained abnormality rates exceeds a predetermined value, the object to be tested is determined to be abnormal. It has 9 steps.

The present invention is a condition monitoring device that, in other aspects, diagnoses an object under test using any of the above methods.

The condition monitoring method of the present invention can discriminate an abnormality from measurement data at an earlier stage in a monitored device affected by operating conditions and noise, and can improve the accuracy of the condition monitoring system.

It is a block diagram which shows the structure of the state monitoring apparatus which concerns on this embodiment. It is a block diagram which shows the detail of a data calculation part. It is a conceptual diagram which shows the relationship between the measurement data and segment data. It is a figure for demonstrating the feature vector. It is a figure for demonstrating the basic concept of OC-SVM. It is a flowchart for demonstrating the process performed by the data acquisition part of FIG. It is a flowchart for demonstrating the process performed by the learning part of FIG. It is a flowchart for demonstrating the process performed by the abnormality degree calculation part of FIG. It is a flowchart for demonstrating the process performed by the discriminant part of FIG. It is a graph which shows the effect of calculating an abnormality rate average. It is a conceptual diagram for demonstrating that the feature vector is obtained for each segment set. It is a figure which shows the abnormality determination result.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings below, the same or corresponding parts are given the same reference numbers, and the explanations are not repeated.

[Basic configuration of condition monitoring device]
FIG. 1 is a block diagram showing a configuration of a condition monitoring device according to the present embodiment. With reference to FIG. 1, the condition monitoring device 100 receives a signal from the vibration sensor 20 installed in the test device 10 to monitor the state of the test device 10 and detect an abnormality. The device under test 10 is equipment including a rotating device installed in, for example, a factory or a power plant, and the vibration sensor 20 can detect abnormal vibration generated during rotation. In the present embodiment, vibration is illustrated as a monitoring target, but a detection signal other than the vibration sensor may be used as long as it is an output signal that can confirm the operating status of the equipment. For example, a sensor that detects sound, temperature, load torque, motor power, and the like may be used instead of the vibration sensor 20.

The condition monitoring device 100 includes an A / D converter 110, a data acquisition unit 120, a storage device 130, a data calculation unit 140, and a display unit 150.

The A / D converter 110 receives the output signal of the vibration sensor 20. The data acquisition unit 120 receives a digital signal from the A / D converter 110, performs filtering processing, and records the measurement data in the storage device 130. The data calculation unit 140 reads the measurement data measured at the time of normal from the storage device 130 to create an abnormality discrimination threshold value for discriminating an abnormality, or uses the abnormality discrimination threshold value to measure the measurement during the test. From the data, it is determined whether or not there is an abnormality in the device under test 10. When the data calculation unit 140 determines the presence or absence of an abnormality, the data calculation unit 140 causes the display unit 150 to display the result.

FIG. 2 is a block diagram showing details of the data calculation unit. The data calculation unit 140 includes a learning unit 142, a threshold value storage unit 144, an abnormality degree calculation unit 146, and a discrimination unit 148.

The learning unit 142 classifies the classification boundary, which is the boundary for classifying normal and abnormal based on the data (normal data) acquired from the storage device 130 by the test device 10 in the normal state (initial state, etc.). An abnormality determination threshold value for determining the degree of abnormality corresponding to the distance from the boundary is generated and stored in the threshold value storage unit 144.

The abnormality degree calculation unit 146 applies the classification boundary to the data (test data) acquired from the storage device 130 at the time of diagnosis of the test device 10, calculates the abnormality degree corresponding to the distance from the classification boundary, and causes the discrimination unit 148 to calculate the abnormality degree. send.

The determination unit 148 determines the abnormality of the device under test 10 based on the result of comparing the degree of abnormality and the abnormality determination threshold value.

One of the features of the condition monitoring device 100 of the present embodiment is that when processing the measurement data, the entire measurement data is not processed as a whole, but the entire measurement data is divided into a plurality of segments and processed for each segment. .. The segments will be described below.

FIG. 3 is a conceptual diagram showing the relationship between the measurement data and the segment data. With reference to FIG. 3, the measurement data is data having a time length of T1, and after the output signal from the vibration sensor 20 is A / D converted by the A / D converter 110 and filtered by the data acquisition unit 120, This is the data stored in the storage device 130.

The measurement data stored in the storage device 130 includes training data and test data. The learning data is measurement data acquired when the device under test 10 is known to be normal (for example, in the initial state). The test data is measurement data acquired when it is desired to determine the normality / abnormality of the device under test 10. The training data determines the classification boundaries and abnormality discrimination thresholds described later. As a result of performing predetermined processing on the test data using the classification boundary and the abnormality discrimination threshold value, the normality / abnormality of the test device 10 is determined.

The measurement data of the time length T1 is divided into segments of the time length T2 shorter than the time length T1 as shown in FIG. 3 when processed by the learning unit 142 and the abnormality degree calculation unit 146. For example, when the time length T1 is 20 seconds, the time length T2 can be set to 0.2 seconds. In this case, one measurement data is divided into 100 segment data. After the measurement data is divided, the learning unit 142 and the abnormality degree calculation unit 146 further create a feature amount vector for each segment.

FIG. 4 is a diagram for explaining the feature quantity vector. FIG. 4 shows an example in which the measurement data is divided into m segments and the feature amount is n.

The features are, for example, effective value (OA), maximum value (Max), crest factor, kurtosis, skewness, and after these signal processing (FFT processing, when the measurement data is vibration). It can be a value of kurtosis processing). The feature amount vector treats a plurality of feature amounts as a set of vectors. These feature vectors are used for abnormality determination. For one measurement data, m feature quantity vectors 1 to m are created.

If the extraction of the feature amount and the creation of the feature amount vector are processed together with the entire measurement data, there is a risk that the entire measurement data cannot be used for diagnosis when a sudden abnormality occurs. Therefore, in the present embodiment, the measurement data is divided into segments, and the feature amount is extracted and the feature amount vector is calculated for each segment. For example, when a rotating device is monitored by a vibration sensor, a temporary impact such as dropping a tool or the like during acquisition of measurement data may be detected by the vibration sensor as a sudden vibration. If the features are extracted by dividing them into segments, the correct features can be extracted at times other than the time of sudden abnormality even in such a case, and by comparing the features for each segment, the segment corresponding to the sudden abnormality can be extracted. It is also possible to remove and evaluate.

The degree of abnormality 1 to m is calculated for the feature vector 1 to m based on the classification boundary. The classification boundary is an index for performing anomaly discrimination used in a known anomaly detection method (One Class Support Vector Machine: OC-SVM).

FIG. 5 is a diagram for explaining the basic concept of OC-SVM. In FIG. 5, circled “○” indicates learning data acquired for learning the normal state when the device under test 10 is known to be normal, and square and triangular marks “◯”. □, △ ”indicates the test data to be diagnosed. Of the test data, "□" is data indicating abnormality, and "Δ" is data indicating normality.

For example, as shown in Fig. 5 (a) on the left, on a two-dimensional scatter plot when there are two features, the training data and the test data may not have a boundary line that can classify normal / abnormal. Think. Since useful features differ depending on the diagnosis target and operating conditions, an appropriate feature is selected. By mapping each training data and test data to a multidimensional feature space containing appropriate features, it becomes possible to generate a classification interface that can classify normal / abnormal. For each training data and test data, the degree of anomaly, which is the distance from the classification boundary, can be calculated. The degree of abnormality is zero on the classification boundary, the degree of abnormality is a negative (-) value on the normal side of the classification boundary, and the degree of abnormality is a positive (+) value on the abnormal side.

Such a method is called machine learning by OC-SVM, and it is possible to convert many features into one index (abnormality) and evaluate it.

The learning unit 142 of FIG. 2 defines the above classification boundary and also determines the abnormality determination threshold value for determining the degree of abnormality of the test data. Further, the abnormality degree calculation unit 146 of FIG. 2 calculates the abnormality degree which is the distance from the classification boundary of each measurement data in the feature space. The discrimination unit 148 of FIG. 2 compares the degree of abnormality with the abnormality discrimination threshold value, calculates the abnormality rate of each measurement data, and outputs the discrimination result.

FIG. 6 is a flowchart for explaining the process performed by the data acquisition unit of FIG. In step S1, the data acquisition unit 120 receives the digitally converted data of the signal including the vibration waveform from the vibration sensor 20, and in step S2, the abnormality to be observed among the low-pass filter, band-pass filter, high-pass filter, and the like. An appropriate filter process is applied to the phenomenon to remove basic noise, and the data is stored in the storage device 130 in step S3.

It should be noted that the data acquisition unit 120 wants to acquire learning data and make a diagnosis while using the test device 10 when it is known that the test device 10 operates normally, such as in the initial state and when the repair is completed. In that case, the test data is automatically acquired at the time specified by the timer or the like.

FIG. 7 is a flowchart for explaining the process performed by the learning unit of FIG. First, the learning unit 142 initializes the count variable j to 1 in step S11. Then, in step S12, D1 learning data is randomly selected from the plurality of measurement data acquired at the normal time. Subsequently, as described with reference to FIG. 3, each learning data is divided into segments (step S13), and the feature amount vector is calculated for each segment (step S14).

Subsequently, in step S15, the learning unit 142 calculates the classification boundary and the abnormality discrimination threshold value using OC-SVM using the D1 × m feature amount vectors included in the selected learning data. Further, when the jth classification boundary and the discrimination threshold value are calculated in step S16, a count variable is added to them and stored.

The count variable j is counted up in step S18 while the count variable j is less than the number of repetitions K1 in step S17 (NO in S16) so that the above processes of S12 to S16 are repeated by the number of repetitions K1. In step S12, the combination of measurement data already selected is not used. When the count variable j reaches the number of repetitions K1 in step S17, the calculation of the K1 classification boundary and the discrimination threshold value is completed, and the process is completed in step S19.

FIG. 8 is a flowchart for explaining the process performed by the abnormality degree calculation unit of FIG. The abnormality degree calculation unit 146 first initializes the count variable j to 1 in step S21. Then, in step S22, D2 test data are randomly selected from the plurality of measurement data acquired at the time of diagnosis. Subsequently, as described with reference to FIG. 3, each test data is divided into segments (step S23), and the feature amount vector is calculated for each segment (step S24).

Subsequently, in step S25, the abnormality degree calculation unit 146 uses the classification boundary of the jth selected learning data generated by the learning unit 142 and held in the threshold storage unit 144 to test data. Calculate the degree of anomaly in each segment of.

The count variable j is counted up in step S27 while the count variable j is less than the number of repetitions K1 in step S26 (NO in S16) so that the above processes of S22 to S25 are repeated by the number of repetitions K1. In step S22, the combination of measurement data already selected as test data is not used.

When the degree of abnormality of each segment having the number of repetitions K1 is calculated in step S26, the process ends in step S28.

FIG. 9 is a flowchart for explaining the process performed by the discriminating unit of FIG. First, in step S31, the determination unit 148 initializes both the count variables i and j to 1. Then, the j-th selected test data (test data j) is targeted (step S32), and the degree of abnormality of the i-th segment (segment i) of the test data is determined by the abnormality of the j-th selected training data. An abnormality determination is made by comparing with the determination threshold value (step S33). While the count variable i is less than the total number of segments D2 × m in step S34 (NO in S34), the count variable i is counted in step S35 so that the above processes of S32 to S35 are repeated by the total number of segments D2 × m. Will be uploaded.

When the degree of abnormality of each segment is calculated for the segments 1 to D2 × m of the test data j in step S34, the determination unit 148 calculates the abnormality rate of the test data j in step S36.

As shown in FIG. 4, the abnormality rate is the total number of segments D2 × m, which is the number of the abnormalities 1 to D2 × m of the segments 1 to D2 × m whose abnormality exceeds the abnormality determination threshold value. It is calculated by dividing by.

The count variable j is counted up in step S38 while the count variable j is less than the number of repetitions K1 in step S37 (NO in S37) so that the above processes of S32 to S36 are repeated by the number of repetitions K1.

When the abnormality rate of each test data is calculated for the test data 1 to K1 in step S37, the determination unit 148 averages the abnormality rates 1 to K1 in step S39 to calculate the abnormality rate average, and in step S40. End the process.

As described above, by dividing the training data and the test data into segments and obtaining the feature quantity vector, measurement data that could not be used as error data in the past can be used. In addition, by repeating the calculation of the average abnormality rate by randomly selecting from a plurality of measurement data and performing the calculation, the abnormality rate is settled and the discrimination result is stabilized.

[Example 1]
A verification experiment was conducted on the condition monitoring method described above. The device under test is a bearing, and an example of condition monitoring of a bearing with artificial damage on the raceway surface is shown.

The vibration acceleration of a bearing having an outer ring raceway of an angular contact ball bearing provided with fine cylindrical holes and rectangular grooves by electric discharge machining was measured when the bearing was operated at a constant speed under a radial load and an axial load. There are five types of discharge hole diameter and groove shape (hereinafter referred to as damage size) shown below. Vibration acceleration was measured 11 times for each damage size. In addition, the testing machine was disassembled and reassembled for each measurement. The operating conditions and measurement conditions are as follows.

<Operating conditions>
Bearing: Angular contact ball bearing (model number 7216: inner diameter 80 mm, outer diameter 140 mm, width 26 mm)
Radial load: 1.3kN
Axial load: 1.3kN
Rotation speed: 1500 rotations / minute Damage size: 0.00mm (normal), φ0.34mm (cylindrical hole), φ0.68mm (cylindrical hole),
φ1.02mm (cylindrical hole), φ1.35mm (cylindrical hole), circumferential direction 2mm x axial direction 10mm x depth 1mm (rectangular groove)
<Measurement conditions>
Measurement data: Vibration acceleration Measurement direction: Vertical direction, horizontal direction, axial direction Data length: 20 seconds Sampling speed: 50kHz
Number of measurements: 11 times / damage size The usefulness of the examples was evaluated using the vibration acceleration data for each damage size obtained above.

<Calculation of feature vector of training data and test data>
Vibration acceleration data obtained in 20 seconds of one measurement is processed every 0.2 seconds (5 rotations of the rotation axis) after frequency filtering (low pass: 20 to 100,000 Hz, band pass: 1000 to 5000 Hz, high pass: 5000 to 20000 Hz). Divided into 100 segments, and the feature quantities (here, effective value OA, maximum value Max, crest factor, sharpness, strain) of the divided measurement data (segment data) in the time domain, frequency domain, and kefrency domain After the calculation and each filter processing at the same time, the feature amount of each region is put together to obtain the feature amount vector (Fig. 3).

<Selection of learning data>
In the normal state, 8 of the 11 measurement data obtained were randomly selected as training data. All feature vector obtained from the selected measurement data is collectively used as learning data.

<Selection of test data>
Three of the 11 feature sets obtained were randomly selected for various damage sizes. As the undamaged data, the remaining 3 data randomly selected as training data were used. The total feature vector obtained from the selected measurement data is used as the test data.

<Creation of classification boundaries and anomaly discrimination thresholds>
A classification boundary is created from the feature vector of the training data using OC-SVM (Fig. 5). Further, the degree of abnormality of the total feature amount vector of the training data is calculated from the created classification boundary, and the abnormality discrimination threshold value is calculated by the following equation (1).
Abnormality discrimination threshold = average value of abnormality degree + 5 x standard deviation of abnormality degree ... (1)
<Calculation of abnormal rate of test data>
The degree of abnormality of each feature vector of the test data is calculated using the classification boundary, and the abnormality rate of the test data is calculated using the following equation (2) (Fig. 4).
Abnormality rate = number of feature vector / total number of feature vectors exceeding the anomaly discrimination threshold value ... (2)
<Diagnosis>
The above process is repeated 10 times, and the average abnormality rate of each damage size is calculated. If the average abnormality rate is 0.5 or more, the test data is regarded as abnormal.

FIG. 10 is a graph showing the effect of calculating the average abnormality rate. FIG. 10A shows the relationship between the abnormality rate and the artificial defect size at each calculation (10 times). FIG. 10B shows the relationship between the average abnormality rate and the artificial defect size when the average abnormality rate is calculated. The artificial defect size of 0be indicates a test piece without defects, and 2be, 4be, 6be, and 8be are φ0.34 mm (cylindrical hole), φ0.68 mm (cylindrical hole), and φ1.02 mm (cylindrical hole), respectively. , Φ1.35 mm (cylindrical hole), and RGe indicates a rectangular groove of 2 mm in the circumferential direction × 10 mm in the axial direction × 1 mm in depth. As can be seen by comparing FIGS. 10 (a) and 10 (b), the variation in each defect size is smaller when the anomaly rate average is calculated. Therefore, it is better to use the anomaly rate average for the abnormality discrimination. It turns out that is easy to stabilize.

[Example 2]
The operating condition and the measuring condition of the second embodiment are the same as those of the first embodiment. A new feature vector obtained by summarizing the feature vectors obtained in Example 1 in units of five consecutive segments (segment sets) in time series was used.

FIG. 11 is a conceptual diagram for explaining that the feature amount vector is obtained for each segment set. Assuming that the time length T1 of the measurement data is 20 seconds, the time length T2 of the segment is 0.2 seconds. Then, the time length T3 of the segment set was set to 2 seconds. In Example 2, the five feature vectors of the five consecutive segments of Example 1 were put together to form a segment set feature vector.

Using the feature vector of this segment set, processing was performed in the same procedure as in Example 1.

[Comparison example]
In the comparative example, an effective value generally used for diagnosis was used as the feature quantity. The calculation method of the comparative example is shown below.

Calculate the vertical effective value at normal time (no damage) and at each damage size. As the effective value, one value is obtained from the entire measurement data.

Eight of the 11 measurement data obtained in the normal state are randomly selected as learning data, and the abnormality discrimination threshold value is calculated from the effective value of the selected measurement data by the following equation (3).
Abnormality discrimination threshold = mean value of effective value + 5 x standard deviation of effective value ... (3)
Three of the measurement data obtained at normal time (no damage) and for each damage size are selected as test data, and the abnormality rate is calculated by the following equation (4) using the effective value of the test data. For normal conditions, use measurement data different from the data for which the abnormality discrimination threshold was created.
Abnormality rate = number of effective values exceeding the abnormality judgment threshold / number of test data ... (4)
The process up to the calculation of the abnormality rate is repeated, and the average abnormality rate of each damage size is calculated. If the average abnormality rate is 0.5 or more, the test data is regarded as abnormal.

[Evaluation]
FIG. 12 is a diagram showing an abnormality determination result. In Examples 1 and 2 and Comparative Example, the result of calculating the average abnormality rate 5 times by changing the training data and the test data is shown in FIG. If the average abnormality rate is 0.5 or more in all 5 calculations, it is marked as ○, if it is 0.5 or more even once, it is marked as △, and if it is not exceeded 0.5 once, it is marked as ×. .. Compared with the comparative example, the abnormality discrimination accuracy was improved in each of Examples 1 and 2.

As described above, the condition monitoring method of the present invention can determine an abnormality from the measurement data at an earlier stage in a rotating machine affected by operating conditions and noise, and can improve the accuracy of the condition monitoring system.

[Action / Effect]
In the present invention, in a rotating machine affected by operating conditions and noise, it is possible to discriminate minute damage that cannot be discriminated in the past.

The condition monitoring method according to the present embodiment includes the first to ninth steps. In the first step (S1), when the object to be tested is normal, a plurality of first measurement data having a first time length are acquired from the sensor installed on the object to be tested at different timings.

In the second step (S1), at the time of diagnosing the object to be tested, a plurality of second measurement data having a first time length are acquired at different timings.

In the third step (S12), a plurality of training data are randomly selected from the plurality of first measurement data. In the fourth step (S22), a plurality of test data are randomly selected from the plurality of second measurement data. In the fifth step (S13, S14), each of the plurality of training data is divided into segment data for each second time length T2 shorter than the first time length T1, and a plurality of features for each segment data after the division. Create a first feature vector containing the quantity. In the sixth step (S15), a classification boundary for classifying normal and abnormal and an abnormality discrimination threshold value are created from a plurality of first feature quantity vectors created for each segment data for the plurality of training data. In the seventh step (S23, S24), each of the plurality of test data is divided into segment data for each second time length, and a second feature amount vector including a plurality of feature amounts for each segment data after the division. To create. In the eighth step (S25, S36), the degree of abnormality, which is the distance from the classification boundary, is calculated for the second feature vector, and the abnormality of the second feature vector is found for each of the plurality of test data. The anomaly rate is calculated in which the number of degrees exceeding the anomaly discrimination threshold is the ratio to the total number of the second feature vector. In the ninth step (S37 to S39), the third to eighth steps are repeatedly executed a plurality of times, and when the average value of the obtained abnormality rates exceeds a predetermined value, the object to be tested is determined to be abnormal.

In the fifth step (S13, S14) and the seventh step (S23, S24), for example, the data measured with a fixed time length T1 is divided by the time length T2 such as a rotation cycle and a cycle, and then the feature quantity vector is calculated. By treating it as a feature set, it becomes possible to evaluate changes in features over time.

The time length T2 is preferably an integral multiple of the rotation cycle and the operation cycle.
The feature quantities in the 5th step (S13, S14) and the 7th step (S23, S24) are, for example, the effective value, the maximum value, and the wave height in the time domain, the frequency domain, and the kurtosis region after the raw measurement data or the bandpass filter. The rate, kurtosis, and skewness can be exemplified.

In the third step (S12), a normal model can be created in consideration of changes in machine characteristics by randomly selecting a feature vector from long-term output signals including setup and stop / restart. It will be possible.

In the fourth step (S22), as in the third step (S12), changes in machine characteristics are taken into consideration by randomly selecting feature vector from long-term data including setup and stop / restart. It becomes possible to create test data.

In addition to SVM, random forest, logistic regression, decision tree, and neural network can be exemplified as methods used for learning a normal model and creating an abnormality discrimination index and an abnormality discrimination threshold value.

In the ninth step (S37 to S39), the normal model and test data are changed, the abnormality rate is calculated a plurality of times, and the abnormality rate is averaged to prevent erroneous evaluation in a machine whose operating condition fluctuates.

The output signal of the sensor used for condition monitoring may be any output signal that can confirm the operating status of the rotating machine and equipment, and vibration, sound, temperature, load torque, and motor power can be exemplified.

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

10 Tested device, 20 Vibration sensor, 100 Condition monitoring device, 110 A / D converter, 120 Data acquisition unit, 130 Storage device, 140 Data calculation unit, 142 Learning unit, 144 Value storage unit, 146 Abnormality calculation unit, 148 Discrimination unit, 150 display unit.

Claims (3)

When the rotating machine , which is the object to be tested, is normal, the first step of acquiring a plurality of first measurement data having a length of one time from the sensor installed on the object to be tested at different timings.
At the time of diagnosing the object to be tested, a second step of acquiring a plurality of second measurement data having the first time length at different timings, and
A third step of randomly selecting a plurality of training data from the plurality of first measurement data, and
A fourth step of randomly selecting a plurality of test data from the plurality of second measurement data, and
Each of the plurality of training data is divided into segment data for each second time length shorter than the first time length , and the effective value, maximum value, and wave height, which are a plurality of feature quantities , are obtained for each segment data after the division. The fifth step of creating the first feature vector including the rate, kurtosis, and skewness , and
A sixth step of creating a classification boundary for classifying normal and abnormal and an abnormality discrimination threshold value from a plurality of first feature vector created for each segment data of the plurality of training data.
Each of the plurality of test data is divided into segment data for each second time length , and the effective value, maximum value, crest factor, kurtosis, and skewness, which are a plurality of feature quantities , are obtained for each segment data after the division. A seventh step of creating a second feature vector containing
The degree of abnormality, which is the distance from the classification boundary, is calculated for the second feature amount vector, and the degree of abnormality of the second feature amount vector determines the abnormality for each of the plurality of test data. The eighth step of calculating the anomaly rate, in which the number exceeding the threshold value is the ratio to the total number of the second feature vector,
Condition monitoring including the ninth step of repeatedly executing the third to eighth steps a plurality of times and determining the object to be tested as an abnormality when the average value of the obtained abnormality rates exceeds a predetermined value. Method. When the rotating machine , which is the object to be tested, is normal, the first step of acquiring a plurality of first measurement data having a length of one time from the sensor installed on the object to be tested at different timings.
At the time of diagnosing the object to be tested, a second step of acquiring a plurality of second measurement data having the first time length at different timings, and
A third step of randomly selecting a plurality of training data from the plurality of first measurement data, and
A fourth step of randomly selecting a plurality of test data from the plurality of second measurement data, and
Each of the plurality of training data is divided into segment data for each second time length shorter than the first time length , and the effective value, maximum value, and wave height, which are a plurality of feature quantities , are obtained for each segment data after the division. A fifth step of creating a first feature vector including rate, kurtosis, and skewness , and collecting the first feature vectors of a plurality of consecutive segment data to create a third feature vector,
A sixth step of creating a classification boundary for classifying normal and abnormal and an abnormality discrimination threshold value from a plurality of third feature vector created for each of the continuous plurality of segment data for the plurality of training data. When,
Each of the plurality of test data is divided into segment data for each second time length , and the effective value, maximum value, crest factor, sharpness, and skewness, which are a plurality of feature quantities , are obtained for each segment data after the division. A seventh step of creating a second feature vector including the above, and combining the second feature vectors of a plurality of consecutive segment data to create a fourth feature vector.
The degree of abnormality, which is the distance from the classification boundary, is calculated for the fourth feature amount vector, and the degree of abnormality of the fourth feature amount vector determines the abnormality for each of the plurality of test data. The eighth step of calculating the anomaly rate, in which the number exceeding the threshold value is the ratio to the total number of the fourth feature vector,
Condition monitoring including the ninth step of repeatedly executing the third to eighth steps a plurality of times and determining the object to be tested as an abnormality when the average value of the obtained abnormality rates exceeds a predetermined value. Method. A condition monitoring device for diagnosing the object to be tested by using the method according to claim 1 or 2.

Images