JP7083293B2 - Status monitoring method and status monitoring device - Google Patents

Description

The present invention relates to a condition monitoring method and a condition monitoring device for a device under test such as a rotary machine.

Conventionally, in a test object such as a rotary machine, the state of the test object is monitored based on the measurement data of physical quantities measured by various sensors. Specifically, a feature amount indicating the characteristic of the waveform shown by the measurement data is calculated, and the presence or absence of abnormality of the test object is determined using the calculated feature amount.

The accuracy of determining the presence or absence of an abnormality depends on the sensitivity of the feature amount to the abnormality. Therefore, condition monitoring methods using various features have been developed. For example, in Japanese Patent Application Laid-Open No. 10-274558 (Patent Document 1), the presence or absence of an abnormality is determined by using a set of a frequency at which a peak occurs and a time interval at which the peak occurs, which is obtained from a time change in the spectrum of waveform data. The method of diagnosing abnormalities is disclosed. Japanese Patent Application Laid-Open No. 2006-300895 (Patent Document 2) determines the presence or absence of equipment abnormality by using the minimum distance between each neuron in the clustering map and the neuron corresponding to the frequency component extracted from the signal during operation of the equipment. The equipment monitoring method to be performed is disclosed.

Japanese Unexamined Patent Publication No. 10-274558 Japanese Unexamined Patent Publication No. 2006-300895

However, it is desired to further improve the accuracy of determining the presence or absence of abnormality of the object to be tested.

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 with high accuracy in determining the presence or absence of an abnormality in a test object.

The present invention is a condition monitoring method for monitoring the condition of an object to be tested by using measurement data acquired from a sensor installed on the object to be tested. The condition monitoring method includes first to fourth steps. In the first step, the measurement data is filtered using each of a plurality of bandpass filters having different center frequencies, thereby generating a plurality of first waveforms corresponding to the plurality of bandpass filters. do. In the second step, the first feature amount is calculated for each of the plurality of first waveforms. In the third step, a graph having the center frequency as the horizontal axis and the first feature amount as the vertical axis is generated for each of the plurality of bandpass filters by using the center frequency of the bandpass filter and the bandpass filter. The second feature amount indicating the feature of the second waveform obtained by plotting the calculated first feature amount with respect to the calculated time waveform is calculated. In the fourth step, the test object is tested based on the second feature amount calculated by performing the first to third steps on the test data obtained from the sensor at the time of diagnosing the test object among the measurement data. Determine if an object is abnormal.

Preferably, the plurality of first waveforms is a time waveform. The first feature amount is a first waveform corresponding to the first feature amount among a plurality of first waveforms, a third waveform obtained by Fourier transforming the corresponding first waveform, and a Fourier transform of the third waveform. The characteristics of any one of the waveforms with the fourth waveform obtained in the above are shown.

Preferably, the plurality of first waveforms is a time waveform. In the second step, the first waveform corresponding to the first feature amount in the plurality of first waveforms, the third waveform obtained by Fourier transforming the corresponding first waveform, and the third waveform are Fourier transformed. A step of calculating a third feature amount indicating the characteristics of the waveform for each of at least two waveforms of the obtained fourth waveform, and a third feature amount calculated for each of the at least two waveforms. 1 Includes a step of calculating a feature amount.

Preferably, the state monitoring method uses the second feature amount obtained by performing the first to third steps on the normal data obtained from the sensor when the test object is normal among the measurement data as the learning data. By performing machine learning, a fifth step of determining an algorithm for determining whether or not the object to be tested is abnormal is further provided. In the fourth step, it is determined whether or not the test object is abnormal according to an algorithm.

Preferably, the second feature amount is any one of the effective value, the maximum value, the crest rate, the kurtosis, the skewness, the coefficient of variation and the contribution rate of the second waveform.

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 improve the accuracy of determining the presence or absence of abnormality in the device under test.

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 flowchart for demonstrating the process performed by the data acquisition part of FIG. It is a figure which shows an example of the measurement data acquired by the data acquisition part of FIG. It is a figure which shows another example of the measurement data acquired by the data acquisition part of FIG. It is a flowchart for demonstrating the processing performed by the filter processing unit and the feature amount calculation unit of FIG. The three time waveforms generated from the measurement data of FIG. 4 are shown. The three time waveforms generated from the measurement data of FIG. 5 are shown. It is a figure which shows the example of the center frequency waveform. 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 determination part of FIG. It is a figure which shows the change of the degree of abnormality between an Example and a comparative example. It is a figure which shows the t-test result of the 2nd feature quantity calculated by the data calculation method of an Example in the verification experiment 2. It is a figure which shows the t-examination result of the comparative feature quantity calculated by the data calculation method of the comparative example. It is a figure which shows an example of the 1st feature amount of the modification. It is a figure which shows the t-test result of the 2nd feature amount in the verification experiment 3. It is a figure which shows the t-test result of the 2nd feature quantity in the verification experiment 4.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following drawings, the same or corresponding parts will be given the same reference number, and the explanation will not be 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 vibration acceleration signal from the vibration sensor 20 installed in the test device 10, monitors the state of the test device 10, and determines the presence or absence of an abnormality. The device under test 10 is equipment including a rotating machine installed in, for example, a factory or a power plant, and the vibration sensor 20 can detect abnormal vibration generated during rotation. In this embodiment, the vibration sensor 20 that outputs the vibration acceleration signal is exemplified, 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 outputs signals indicating physical quantities such as speed, displacement, sound, AE (Acoustic Emission), temperature, load torque, and motor power 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 vibration acceleration signal of the vibration sensor 20, converts it into a digital format signal, and outputs the signal. The data acquisition unit 120 receives a digital vibration acceleration signal from the A / D converter 110, performs noise removal processing, and generates measurement data indicating a change in vibration acceleration for a predetermined time (for example, 20 seconds). The data acquisition unit 120 records the generated measurement data in the storage device 130. The data calculation unit 140 reads the measurement data (normal data) measured at the normal time from the storage device 130, determines an algorithm for determining the presence / absence of an abnormality, and measures at the time of the test using the algorithm. From the measurement data (test 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 determination result.

FIG. 2 is a block diagram showing details of the data calculation unit. The data calculation unit 140 includes a filter processing unit 141, a feature amount calculation unit 142, a learning unit 143, a threshold value storage unit 144, and a determination unit 146.

The filter processing unit 141 performs filter processing on the measurement data stored in the storage device 130 using each of a plurality of bandpass filters (BPF (Band-pass Filter)) having different center frequencies. , Generates a plurality of time waveforms (first waveform) corresponding to each of the plurality of bandpass filters.

The feature amount calculation unit 142 calculates the first feature amount for each of the plurality of time waveforms generated by the filter processing unit 141. Further, the feature amount calculation unit 142 obtains a waveform (hereinafter referred to as “center frequency waveform”) (second waveform) obtained by arranging the first feature amount in ascending order of the center frequency of the bandpass filter for each measurement data. Generate. The feature amount calculation unit 142 calculates a second feature amount indicating the feature of the center frequency waveform corresponding to the measurement data for each measurement data.

The learning unit 143 performs machine learning using a plurality of second feature quantities calculated for each of the plurality of measurement data (normal data) acquired when the device 10 under test is normal (initial state, etc.) as learning data. Thereby, an algorithm for determining whether or not the device under test 10 is abnormal is determined. The learning unit 143 stores the determined algorithm in the threshold value storage unit 144.

The determination unit 146 uses the second feature amount calculated for the measurement data (test data) obtained at the time of diagnosis of the test device 10 according to the algorithm stored in the threshold storage unit 144, and the test device 10 is used. Judge the presence or absence of an abnormality.

[Processing of data acquisition unit]
FIG. 3 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 vibration acceleration signal in the digital format and generates data indicating the change in the vibration acceleration for a predetermined time (for example, 20 seconds). In step S2, the data acquisition unit 120 applies an appropriate filter process among the low-pass filter, band-pass filter, high-pass filter, and the like to the generated data to generate measurement data from which basic noise is removed. In step S3, the data acquisition unit 120 stores the generated measurement data in the storage device 130.

The data acquisition unit 120 stores the measurement data acquired when it is known to operate normally, such as the initial state of the test device 10 and the completion of repair, as normal data in the storage device 130. The data acquisition unit 120 stores the measurement data acquired when it is desired to perform a diagnosis while the test apparatus 10 is in use as test data in the storage apparatus 130. The data acquisition unit 120 automatically acquires test data at a time (for example, every 2 hours) specified by a timer or the like.

FIG. 4 is a diagram showing an example of measurement data acquired by the data acquisition unit 120. FIG. 5 is a diagram showing another example of the measurement data acquired by the data acquisition unit 120. 4 and 5 show changes in vibration acceleration over time when the device under test 10 is a bearing. FIG. 4 shows measurement data when a cylindrical damage having a diameter of 1.35 mm is artificially provided on the raceway surface of the bearing, and FIG. 5 shows artificially larger damage on the raceway surface of the bearing. The measurement data when it is provided in the above is shown.

[Processing of filter processing unit and feature amount calculation unit]
FIG. 6 is a flowchart for explaining the processing performed by the filter processing unit 141 and the feature amount calculation unit 142 of FIG.

In step S11, the filter processing unit 141 performs filter processing on the measurement data using each of the plurality of bandpass filters having different center frequencies. The bandwidth of the plurality of bandpass filters is not particularly limited, and may be constant (for example, 1 kHz) or may be different from each other. The center frequencies of the plurality of bandpass filters may be fixed intervals (for example, 1 kHz) or irregular intervals. Further, the band of one bandpass filter in the plurality of bandpass filters does not have to overlap with the band of any of the remaining bandpass filters, and overlaps with the band of any of the remaining bandpass filters. You may. The filter processing unit 141 generates a plurality of time waveforms corresponding to the plurality of bandpass filters.

FIG. 7 shows three time waveforms generated from the measurement data of FIG. FIG. 8 shows three time waveforms generated from the measurement data of FIG. 7 and 8 (a) show time waveforms generated by using a bandpass filter having a pass band of 0 to 1 kHz. 7 and 8 (b) show time waveforms generated by using a bandpass filter having a pass band of 1 to 2 kHz. 7 and 8 (c) show time waveforms generated by using a bandpass filter having a pass band of 9 to 10 kHz. As shown in FIGS. 7 and 8, the time waveform differs depending on the frequency band.

Returning to FIG. 6, next, in step S12, the feature amount calculation unit 142 calculates the first feature amount for each of the plurality of time waveforms. Specifically, the first feature amount is a time region feature amount showing the characteristics of the corresponding time waveform and a frequency region feature amount showing the characteristics of the frequency waveform (third waveform) obtained by Fourier transforming the time waveform. And the Kefrensi region feature amount which shows the characteristics of the Kefrenshi waveform (fourth waveform) obtained by Fourier transforming the frequency waveform as a time waveform. The first feature amount is, for example, the effective value, maximum value, crest factor, kurtosis, skewness, peak-to-peak value, coefficient of variation, standard deviation, variance, correlation coefficient, contribution rate, and waveform after envelope processing. Is the effective value of. Alternatively, the first feature quantity may be a common logarithm of any of these parameters. The correlation coefficient is a coefficient indicating the degree of the relationship (for example, a linear relationship) of the distribution law of two random variables (here, time, frequency or kefrency and acceleration). The contribution rate is the squared value of the correlation coefficient.

The feature amount calculation unit 142 may calculate one type of first feature amount or a plurality of types of first feature amount for one time waveform. For example, the feature amount calculation unit 142 may calculate the effective value and the kurtosis for each of the three waveforms of the time waveform, the frequency waveform, and the kurtosis waveform. In this case, the feature amount calculation unit 142 calculates 3 × 2 = 6 types of first feature amounts for one time waveform.

Next, in step S13, the feature amount calculation unit 142 generates a center frequency waveform having the center frequency of the bandpass filter as a variable. That is, the feature amount calculation unit 142 displays, for each of the plurality of bandpass filters, the center frequency of the bandpass filter and the center frequency of the bandpass filter in a graph having the center frequency of the bandpass filter as the horizontal axis and the first feature amount as the vertical axis. A center frequency waveform is generated by plotting points indicating the calculated first feature amount with respect to the time waveform corresponding to the bandpass filter.

FIG. 9 is a diagram showing an example of a center frequency waveform. FIG. 9 shows a center frequency waveform when the first feature amount is the maximum value of the frequency waveform. In FIG. 9, the center frequency waveform described as “damage size 1.35 mm” indicates a waveform generated from the measurement data of FIG. 4, and the center frequency waveform described as “no damage” is the device under test 10. The waveform generated from the measurement data when the bearing is not damaged is shown. As shown in FIG. 9, the shape of the center frequency waveform when there is no damage and the shape of the center frequency waveform when there is damage are different. This means that the shape of the center frequency waveform is affected by the presence or absence of damage.

When the feature amount calculation unit 142 calculates a plurality of types of first feature amounts, the feature amount calculation unit 142 generates a plurality of center frequency waveforms corresponding to each of the plurality of types of first feature amounts.

Returning to FIG. 6, next, in step S14, the feature amount calculation unit 142 calculates the second feature amount indicating the feature of the center frequency waveform. The second feature amount is, for example, the effective value, maximum value, crest factor, kurtosis, skewness, peak-to-peak value, coefficient of variation, standard deviation, variance, correlation coefficient and contribution rate, and envelope processing of the central frequency waveform. It is one of the effective values of the later center frequency waveform.

The correlation coefficient is a coefficient indicating the degree of the relationship (for example, a linear relationship) of the distribution law of two random variables (here, the center frequency and the first feature amount). The contribution rate is the squared value of the correlation coefficient. A sample correlation coefficient can be used as the correlation coefficient. The sample correlation coefficient ρ is the first feature corresponding to the time waveform generated by using the bandpass filter by setting the value of the center frequency of the bandpass filter to _xi (i = 1, 2, ..., N). When the value of the quantity is y _i (i = 1, 2, ..., N), it is calculated by the following equation 1. In Equation 1, x _ave is the average value of x _i (i = 1, 2, ..., N), and y _ave is the average of y _i (i = 1, 2, ..., N). The value.

If it is known in advance that the first feature amount is likely to change significantly in a specific range of the center frequency, the correlation coefficient is calculated using the value of the first feature amount in which the center frequency is within the specific range. May be done.

The feature amount calculation unit 142 may calculate one type of second feature amount or a plurality of types of second feature amount for one center frequency waveform. For example, the feature amount calculation unit 142 may calculate seven types of second feature amounts of the effective value, the maximum value, the crest rate, the kurtosis, the skewness, the coefficient of variation, and the contribution rate with respect to the center frequency waveform. ..

[Processing of learning department]
FIG. 10 is a flowchart for explaining the process performed by the learning unit 143 of FIG. First, in step S21, the learning unit 143 acquires the second feature amount calculated from each of the plurality of normal data measured during the first normal period in which the device under test 10 is normal.

Next, in step S22, the learning unit 143 performs machine learning using the second feature amount acquired in step S21 as learning data, and determines an arithmetic expression for calculating the degree of abnormality for determining the presence or absence of an abnormality. Specifically, the learning unit 143 uses a known machine learning method (One Class Support Vector Machine: OC-SVM) to determine the classification boundary between abnormal and normal from the learning data. Then, the learning unit 143 sets the distance from the determined classification boundary as the degree of abnormality, and determines the arithmetic expression for calculating the degree of abnormality.

As described above, the feature amount calculation unit 142 can calculate the first feature amount of m types (m is an integer of 1 or more) for one measurement data, and can be used as one type of first feature amount. On the other hand, it is possible to calculate the second feature amount of n types (n is an integer of 1 or more). Therefore, the type of the second feature amount output from the feature amount calculation unit 142 for one measurement data is m × n types. The learning unit 143 generates a feature amount vector which is a component of each of m × n kinds of second feature amounts calculated for one normal data. Using OC-SVM, the learning unit 143 determines the classification boundary from the same number of feature quantity vectors as the number of normal data, and determines the arithmetic expression for calculating the degree of abnormality, which is the distance from the determined classification boundary. do.

Next, in step S23, the learning unit 143 acquires the second feature amount calculated from each of the plurality of normal data measured during the second normal period in which the device under test 10 is normal. The second normal period is the period after the first normal period. In step S24, the learning unit 143 calculates the degree of abnormality for each normal data using the second feature amount acquired in step S23 according to the arithmetic expression determined in step S22. In step S25, the learning unit 143 determines the abnormality determination threshold value based on the variation in the degree of abnormality calculated for each normal data. For example, the learning unit 143 determines a value obtained by adding a value that is a constant multiple (for example, 3 times) of the standard deviation of the abnormality degree to the average value of the abnormality degree as an abnormality determination threshold value. In step S26, the learning unit 143 uses the calculation formula of the degree of abnormality determined in step S22 and the abnormality determination threshold value determined in step S25 as an algorithm for determining whether or not the device under test 10 is abnormal. It is stored in the threshold value storage unit 144.

Note that step S23 is omitted, and in step S24, the learning unit 143 calculates the degree of abnormality for each normal data using the second feature amount acquired in step S21 according to the arithmetic expression determined in step S22. May be good.

[Processing of judgment unit]
FIG. 11 is a flowchart for explaining the process performed by the determination unit 146 of FIG. First, in step S31, the determination unit 146 acquires the second feature amount calculated from the measurement data (test data) measured at the time of diagnosis of the device under test 10. In step S32, the determination unit 146 calculates the degree of abnormality using the second feature amount acquired in step S31 according to the calculation formula stored in the threshold value storage unit 144. In step S33, the determination unit 146 compares the degree of abnormality calculated in step S32 with the abnormality determination threshold value stored in the threshold value storage unit 144, and based on the comparison result, the abnormality of the device under test 10 is found. Determine the presence or absence.

[Experiment 1 for verifying the effect of this embodiment]
A verification experiment was conducted to confirm that the degree of abnormality calculated by the above-mentioned condition monitoring method is superior to the degree of abnormality calculated by a known method.

A new bearing that was not damaged was used as the equipment to be tested, and the bearing was continuously operated until it became unusable due to an abnormality. The measurement data of the vibration acceleration measured by the vibration sensor installed on the bearing was acquired every two hours. 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 rpm Lubrication method: Grease <Measurement conditions>
Measurement data: Vibration acceleration data Length: 20 seconds Sampling speed: 50kHz.

For the measurement data group acquired under such operating conditions and measurement conditions, the change in the degree of abnormality calculated according to the data calculation method of the embodiment and the change in the degree of abnormality calculated according to the data calculation method of the comparative example. Was compared.

The data calculation method in the examples is as follows.
Processing method by the filter processing unit 141: The bandwidth of each bandpass filter is 1 kHz, the interval between the center frequencies of a plurality of bandpass filters is 1 kHz, and the frequency range is 0 to 20 kHz, and 20 time waveforms are generated from one measurement data. ..
First feature amount calculated by the feature amount calculation unit 142 for one measurement data: 6 types (effective value and kurtosis of each of time waveform, frequency waveform and kurtosis waveform).
-Second feature amount calculated by the feature amount calculation unit 142 for one center frequency waveform: 7 types (effective value, maximum value, crest factor, kurtosis, skewness, coefficient of variation, contribution factor). The contribution rate is the square of the correlation coefficient between the center frequency and the first feature amount in the range of the center frequency of 1.5 to 9.5 kHz.
-Machine learning method of learning unit 143: OS-SVM.

In the data calculation method of the embodiment, 6 × 7 = 42 types of second feature quantities calculated from each of the initial measurement data (normal data) from the first to the predetermined number in the above measurement data group are used as components. The calculation formula of the degree of anomaly was determined by performing machine learning using the feature quantity vector to be used as training data.

On the other hand, the data calculation method of the comparative example is as follows.
-Characteristics to be used: Each of the time waveform shown in the measurement data, the frequency waveform obtained by Fourier transforming the time waveform, and the kefrency waveform obtained by Fourier transforming the frequency waveform as a time waveform. A total of 15 types of effective value, maximum value, wave height rate, sharpness, and distortion.
-Machine learning method: OS-SVM.

In the data calculation method of the comparative example, a feature quantity vector having the above 15 types of feature quantities calculated from each of the initial measurement data (normal data) from the first to the predetermined th in the above measurement data group is used as a component. By performing machine learning as training data, the calculation formula of the degree of abnormality was determined.

FIG. 12 is a diagram showing changes in the degree of abnormality between the examples and the comparative examples. As shown in FIG. 12, although the degree of abnormality is calculated from the same measurement data group, the time point at which the degree of abnormality of the examples begins to increase is higher than the time point at which the degree of abnormality of the comparative example begins to increase. About 3 months early. That is, it can be seen that the degree of abnormality calculated by the embodiment is a value that reflects a slight abnormality of the bearing. As described above, it was confirmed that the presence or absence of an abnormality in the bearing can be accurately determined by using the degree of abnormality calculated in the examples.

[Verification experiment 2 of the effect of this embodiment]
Next, a verification experiment was conducted to verify that the sensitivity of the second feature amount used in the above-mentioned condition monitoring method to the abnormality of the device under test is excellent.

As the test device, a bearing without damage and a bearing with artificial damage were prepared. The artificial damage is a minute cylindrical hole provided by electric discharge machining in the outer ring raceway of the bearing. The vibration acceleration when these bearings were operated at a constant rotational speed under a radial load and an axial load was measured. The rotation speed was set to 3 conditions. There are five types of discharge hole diameters (hereinafter, damage size) shown below (damage size 0.00 mm indicates undamaged bearings). Vibration acceleration was measured 36 times for each damage size. 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: 1000 rpm, 1500 rpm, 2000 rpm Lubrication method: Circulation lubrication Damage size: 0.00 mm (no damage), φ0.34 mm, φ0.68 mm, φ1.02 mm, φ1.35 mm
<Measurement conditions>
Measurement data: Vibration acceleration data Length: 20 seconds Sampling speed: 50kHz
Number of measurements: 36 times.

The data calculation method in the examples is as follows.
Processing method by the filter processing unit 141: The bandwidth of each bandpass filter is 1 kHz, the interval between the center frequencies of a plurality of bandpass filters is 1 kHz, and the frequency range is 0 to 20 kHz, and 20 time waveforms are generated from one measurement data. ..
First feature amount calculated by the feature amount calculation unit 142 for one measurement data: 6 types (effective value and kurtosis of each of time waveform, frequency waveform and kurtosis waveform).
-Second feature amount calculated by the feature amount calculation unit 142 for one center frequency waveform: 7 types (effective value, maximum value, crest factor, kurtosis, skewness, coefficient of variation, contribution factor). The contribution rate was the square of the correlation coefficient between the center frequency and the first feature amount in the range of the center frequency of 1.5 to 9.5 kHz.

On the other hand, the data calculation method of the comparative example is as follows.
-Characteristic amount to be used (hereinafter referred to as "comparative feature amount"): The time waveform shown in the measurement data, the frequency waveform obtained by Fourier transforming the time waveform, and the frequency waveform are regarded as the time waveform and Fourier. A total of 15 types of effective value, maximum value, crest factor, sharpness, and distortion of each of the converted Kefrenshi waveforms.

Feature amount calculated from the measurement data corresponding to the bearing without damage (second feature amount in the example, comparative feature amount in the comparative example) and the feature amount calculated from the measurement data corresponding to the bearing with damage (example). Then, the difference from the second feature amount and the comparative feature amount in the comparative example) was evaluated using the t-test.

FIG. 13 is a diagram showing the t-test result of the second feature amount calculated by the data calculation method of the embodiment. FIG. 14 is a diagram showing t-test results of comparative feature quantities calculated by the data calculation method of the comparative example. In FIGS. 13 and 14, the test result is an absolute value of the common logarithmized value. Here, when the value is 10 or more, there is a clear difference between the feature amount calculated from the measurement data corresponding to the bearing without damage and the feature amount calculated from the measurement data corresponding to the bearing with damage. I can say. That is, the feature quantity showing a numerical value of 10 or more has good sensitivity to the abnormality of the bearing. Therefore, by using a feature amount showing a numerical value of 10 or more, the accuracy of determining the presence or absence of an abnormality in the bearing is improved.

As shown in FIGS. 13 and 14, the t-test results of the examples are generally superior to the t-test results of the comparative examples.

For example, under the conditions of a rotation speed of 1000 times / minute and a damage size of φ1.35 mm, the feature quantity having a t-test result of 30 or more is one type in the data calculation method of the comparative example, whereas the data calculation method of the implementation is carried out. Then there were 7 types. Specifically, the "wave height factor", "sharpness", "distortion degree" and "variation coefficient" of the central frequency waveform having the sharpness of the frequency waveform as the first feature amount, and the effective value of the time waveform are the first. The "contribution rate" of the central frequency waveform used as the feature amount, the "contribution rate" of the central frequency waveform using the effective value of the frequency waveform as the first feature amount, and the central frequency using the effective value of the kefrency waveform as the first feature amount. It is the "contribution rate" of the waveform. Among these, the "sharpness" and "variation coefficient" of the central frequency waveform whose first feature amount is the sharpness of the frequency waveform, and the "contribution rate" of the central frequency waveform whose first feature amount is the effective value of the frequency waveform. And, the "contribution rate" of the central frequency waveform having the effective value of the Kefrensi waveform as the first feature amount showed the result of t examination of 40 or more, and had excellent sensitivity to the abnormality of the bearing.

Further, in the data calculation method of the comparative example, even for a bearing having a maximum damage size of φ1.35 mm, there was no comparative feature quantity showing t-examination results of 10 or more for all three rotation speeds. On the other hand, in the data calculation method of the embodiment, for the bearing having the maximum damage size of φ1.35 mm, 6 kinds of second feature quantities showed t-examination result of 10 or more for all three rotation speeds. .. Specifically, the "effective value", "maximum value" and "variation coefficient" of the center frequency waveform having the sharpness of the frequency waveform as the first feature amount, and the center having the effective value of the time waveform as the first feature amount. The "contribution rate" of the frequency waveform, the "contribution rate" of the central frequency waveform whose first feature value is the effective value of the frequency waveform, and the "contribution rate" of the central frequency waveform whose first feature value is the effective value of the kefrency waveform. ". These second features have excellent sensitivity to bearing anomalies regardless of rotational speed. Among these, the "contribution rate" of the center frequency waveform with the effective value of the frequency waveform as the first feature is the examination result 10 for all three rotation speeds even for bearings with a damage size of φ0.6 mm or more. The above is shown.

As described above, it was confirmed that the second feature amount calculated by the data calculation method of the embodiment includes the feature amount showing excellent sensitivity to the abnormality of the bearing.

In addition, among the 42 types of second feature quantities, those showing t-test results of less than 10 regardless of the damage size and the rotation speed are included. However, as described above, there are a plurality of second feature quantities that show better t-test results than the conventional feature quantities. Therefore, it is possible to accurately determine the presence or absence of an abnormality in the bearing by calculating the degree of abnormality using a feature amount vector having 42 types of second feature amounts as components.

[Verification experiment 3 of the effect of this embodiment]
The same verification experiment 3 as the above verification experiment 2 was performed by changing only the points described in the following (a).
(A) Second feature amount calculated by the feature amount calculation unit 142 for one center frequency waveform: 1 type (contribution rate). The contribution rate was the square of the correlation coefficient between the center frequency of the bandpass filter (the range of the center frequency is not limited) and the first feature amount.

FIG. 16 is a diagram showing the t-test result of the second feature amount in the verification experiment 3. The test result shown in FIG. 16 is an absolute value of the common logarithm value, as in FIG. 13. As shown in FIG. 16, the t-test result of the verification experiment 3 is also generally superior to the t-test result of the comparative example (see FIG. 14). That is, the contribution rate had excellent sensitivity to bearing abnormalities without limiting the range of the center frequency of the bandpass filter when determining the contribution rate.

[Verification experiment 4 of the effect of this embodiment]
The same verification experiment 4 as the above verification experiment 2 was performed by changing only the points described in the following (b) and (c).
(B) First feature amount calculated by the feature amount calculation unit 142 for one measurement data: 6 types (effective value and kurtosis common logarithm of each of time waveform, frequency waveform and kurtosis waveform). Since the kurtosis of the time waveform may take a negative value, the common logarithm of the value obtained by subtracting the minimum value of all the data from each data was used.
(C) Second feature amount calculated by the feature amount calculation unit 142 for one center frequency waveform: 1 type (contribution rate). The contribution rate was the square of the correlation coefficient between the center frequency of the bandpass filter (the range of the center frequency is not limited) and the first feature amount.

FIG. 17 is a diagram showing the t-test result of the second feature amount in the verification experiment 4. The test result shown in FIG. 17 is an absolute value of the common logarithm value, as in FIG. 13. As shown in FIG. 17, the t-test result of the verification experiment 4 is also generally superior to the t-test result of the comparative example (see FIG. 14). That is, even when the common logarithmic value of the parameter value was used as the first feature quantity, the contribution rate had excellent sensitivity to the abnormality of the bearing.

[Modification example]
In the above description, the feature amount calculation unit 142 has a time domain feature amount indicating the characteristics of the time waveform generated by using the bandpass filter, and a frequency indicating the characteristics of the frequency waveform obtained by Fourier transforming the time waveform. Either the region feature quantity or the kefrency region feature quantity showing the characteristics of the kefrency waveform obtained by Fourier transforming the frequency waveform was calculated as the first feature quantity. However, in step S12 (see FIG. 6), the feature amount calculation unit 142 calculates at least two feature amounts of the time region feature amount, the frequency region feature amount, and the kefrance region feature amount, and at least two of the feature amounts. The step of calculating the first feature amount from the feature amount may be performed. For example, the feature amount calculation unit 142 establishes at least two correlations between the time region feature amount (third feature amount), the frequency region feature amount (third feature amount), and the kefrance region feature amount (third feature amount). The indicated values (sum, difference, ratio, average) and the like are calculated as the first feature quantity.

FIG. 15 is a diagram showing an example of the first feature amount of the modified example. In FIG. 15, (a) shows the maximum value p of the frequency waveform, which is one of the frequency domain features, and (b) shows the maximum value q of the kefrency waveform, which is one of the kefrency region features. Shown, (c) shows the first feature quantity. As shown in FIG. 15, the feature amount calculation unit 142 has, for each of the plurality of bandpass filters, the maximum value p of the frequency waveform obtained from the time waveform generated by the bandpass filter, and the frequency. The ratio (q / p) of the Kefrensi waveform obtained from the waveform to the maximum value q is calculated as the first feature amount.

In the above description, the learning unit 143 performed machine learning using OC-VSM. However, the learning unit 143 may perform machine learning using a random forest, logistic regression, decision tree, or neural network in addition to SVM.

As shown in FIG. 13, when a specific second feature amount for which the t-test result is excellent (in FIG. 13, the contribution rate when the first feature amount is the “effective value of the frequency waveform”) is known in advance. The feature amount calculation unit 142 may calculate only the specific second feature amount. In this case, the learning unit 143 determines only the abnormality determination threshold value. The determination unit 146 may determine the presence or absence of an abnormality in the test device 10 based on the comparison result between the second feature amount calculated for the test data and the abnormality determination threshold value.

[Action / Effect]
The state monitoring method according to the present embodiment is a method of monitoring the state of the test device 10 by using the measurement data acquired from the vibration sensor 20 installed in the test device (object to be tested) 10. -Provides a fourth step. In the first step (S11), the measurement data is filtered using each of a plurality of bandpass filters having different center frequencies, so that a plurality of time waveforms corresponding to the plurality of bandpass filters are used. (First waveform) is generated. In the second step (S12), the first feature amount is calculated for each of the plurality of time waveforms. In the third step (S13, S14), the center frequency of the bandpass filter and the bandpass of each of the plurality of bandpass filters are shown in a graph having the center frequency as the horizontal axis and the first feature amount as the vertical axis. A second feature amount indicating the characteristics of the center frequency waveform (second waveform) obtained by plotting the first feature amount calculated with respect to the time waveform generated by using the filter is calculated. In the fourth step (S32, S33), among the measurement data, the test data obtained from the sensor at the time of diagnosis of the device under test 10 is based on the second feature amount calculated by performing the first to third steps. Then, it is determined whether or not the device under test 10 is abnormal.

According to the above configuration, it is determined whether or not the test device 10 is abnormal based on the second feature amount having good sensitivity to the abnormality of the test device 10. This makes it possible to improve the accuracy of determining the presence or absence of abnormality in the device under test 10. As a result, the abnormality of the device under test 10 can be detected at an early stage.

When an abnormality occurs in the device under test 10, a change is often seen in a specific frequency band of the measurement data. Therefore, by extracting only the specific frequency band, it becomes easy to determine the presence or absence of an abnormality. Therefore, in the conventional condition monitoring method, in order to improve the accuracy of determining the presence or absence of an abnormality, the frequency band effective for determining the presence or absence of an abnormality has been optimized in advance. However, since the frequency band in which the change is observed differs depending on the content of the abnormality, it is necessary to optimize the frequency band for each of the test devices 10.

According to the above configuration, the second feature quantity indicates the feature of the center frequency waveform. The center frequency waveform is obtained by plotting the center frequency of each of the plurality of bandpass filters and the corresponding first feature amount on a graph having the center frequency as the horizontal axis and the first feature amount as the vertical axis. Even if the specific frequency band in which the change occurs due to the abnormality of the test device 10 is unknown, the change in the center frequency portion included in the specific frequency band in the center frequency waveform can be seen according to the degree of the abnormality. .. Therefore, it is possible to easily determine the presence or absence of an abnormality in the device under test 10 by confirming the second feature amount that reflects the change in the portion.

As the condition monitoring method, machine learning is performed using the second feature amount obtained by performing steps S11 to S14 on the normal data obtained from the vibration sensor 20 when the test device 10 is normal among the measurement data as learning data. By doing so, a fifth step (S22, S24, S25) for determining an algorithm for determining whether or not the device under test 10 is abnormal is further provided. In steps S32 and S33, the determination unit 146 determines whether or not the test device 10 is abnormal according to the algorithm.

As a result, even if there are a plurality of types of the second feature amount, it is possible to determine whether or not it is abnormal according to an algorithm that comprehensively considers these.

The first feature amount is, for example, a corresponding time waveform, a frequency waveform obtained by Fourier transforming the time waveform (third waveform), and a kefrency waveform obtained by Fourier transforming the frequency waveform as a time waveform (third waveform). The characteristics of any one of the waveforms (fourth waveform) are shown.

Alternatively, in step S12, the time waveform corresponding to the first feature amount in the plurality of time waveforms, the frequency waveform obtained by Fourier transforming the corresponding time waveform, and the frequency waveform are regarded as the time waveform and Fourier transformed. A step of calculating a feature amount (third feature amount) indicating the characteristics of the waveform for each of at least two waveforms with the Kefrenshi waveform obtained in the above process, and a calculation for each of the at least two waveforms. It may include a step of calculating the first feature amount from the feature amount.

The second feature amount is, for example, any one of the effective value, the maximum value, the wave height rate, the kurtosis, the skewness, the coefficient of variation, and the contribution rate of the second waveform.

The embodiments disclosed this time should be considered to be exemplary and not restrictive in all respects. The scope of the present invention is shown by the scope of claims rather than the description of the embodiments 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, 141 Filter processing unit, 142 Feature amount calculation unit, 143 Learning unit, 144 Threshold storage unit, 146 judgment unit, 150 display unit.

Claims (5)

It is a condition monitoring method for monitoring the condition of the object to be tested by using the measurement data acquired from the sensor installed on the object to be tested.
A first waveform that generates a plurality of first waveforms corresponding to the plurality of bandpass filters is generated by performing a filtering process on the measured data using each of the plurality of bandpass filters having different center frequencies. Process and
The second step of calculating the first feature amount for each of the plurality of first waveforms, and
In the graph with the center frequency as the horizontal axis and the first feature amount as the vertical axis, for each of the plurality of bandpass filters, the center frequency of the bandpass filter and the time generated by the bandpass filter are used. The third step of calculating the second feature amount indicating the feature of the second waveform obtained by plotting the first feature amount calculated with respect to the waveform, and
Among the measured data, the test data obtained from the sensor at the time of diagnosing the object to be tested is subjected to the first to third steps, and the second feature amount is calculated based on the second feature amount. It is equipped with a fourth step of determining whether or not the object is abnormal.
The second feature amount is a condition monitoring method including at least one of the correlation coefficient and the contribution rate of the second waveform . The plurality of first waveforms are time waveforms, and are
The first feature amount includes a first waveform corresponding to the first feature amount among the plurality of first waveforms, a third waveform obtained by Fourier transforming the corresponding first waveform, and the third feature amount. The state monitoring method according to claim 1, wherein the characteristics of any one of the waveforms and the fourth waveform obtained by Fourier transforming the waveforms are shown. The plurality of first waveforms are time waveforms, and are
The second step is
The first waveform corresponding to the first feature amount in the plurality of first waveforms, the third waveform obtained by Fourier transforming the corresponding first waveform, and the third waveform obtained by Fourier transforming. For each of at least two waveforms with the fourth waveform to be obtained, a step of calculating a third feature amount indicating the characteristics of the waveform, and a step of calculating the third feature amount.
The condition monitoring method according to claim 1, further comprising a step of calculating the first feature amount from the third feature amount calculated for each of the at least two waveforms. Machine learning is performed using the second feature amount obtained by performing the first to third steps on the normal data obtained from the sensor when the object to be tested is normal among the measurement data. By doing so, a fifth step of determining an algorithm for determining whether or not the object to be tested is abnormal is further provided.
The condition monitoring method according to any one of claims 1 to 3, wherein in the fourth step, it is determined whether or not the test object is abnormal according to the algorithm. A condition monitoring device for diagnosing the object to be tested by using the method according to any one of claims 1 to 4 .

Images