WO2019168086A1

WO2019168086A1 - Defect detection system, defect model creation program, and defect detection program

Info

Publication number: WO2019168086A1
Application number: PCT/JP2019/007758
Authority: WO
Inventors: 福井　健一; 正嗣北井
Original assignee: 国立大学法人大阪大学; Ntn株式会社
Priority date: 2018-03-01
Filing date: 2019-02-28
Publication date: 2019-09-06
Also published as: JP2019152979A; JP6950891B2

Abstract

In the present invention, an evaluation signal which is a signal of a rolling bearing is divided into segments, a first evaluation vector is generated using a first feature value selected in advance and a second evaluation vector is generated using a second feature value for each segment, a first evaluated defect degree and a second evaluated defect degree for a first normal model are calculated, and a defect degree of the evaluation signal is calculated from an overall normal model using the first evaluated defect degree and the second evaluated defect degree as input vectors. The defect degree is furthermore compared with a threshold value, and the bearing is determined to be defective when a defect rate, which is the ratio at which the defect degree exceeds a defect detection threshold value from among all the information included in the evaluation signal, exceeds a predetermined defect rate threshold value.

Description

Defect detection system, defect model creation program, and defect detection program

Rolling bearing defect detection system, rolling bearing defect model creation program, and rolling bearing defect, which determines a defect in the rolling bearing based on vibration generated from the rolling bearing when the shaft supported by the rolling bearing is rotated It relates to a detection program.

Rolling bearings are mechanical elements for supporting a shaft that rotates around an axis. Specifically, there are a wide variety of fields in which rolling bearings are used, such as moving bodies such as automobiles, ships, and airplanes, industrial machines such as machine tools and conveyance machines, and large plants such as petrochemical and power generation.

When a defect occurs in a rolling bearing, it not only affects the accuracy and operating efficiency of the rotating machine, but if the defect expands, it may cause damage to the machine itself. Therefore, it becomes a subject to detect the defect of a rolling bearing correctly.

For example, in Patent Document 1, as a defect detection method for a rotating machine equipped with a rolling bearing, a normal state model of a physical quantity such as vibration obtained from the rotating machine is created, and the newly measured physical quantity is separated from the normal state. A method for discriminating defects is disclosed. Patent Document 2 discloses a method for identifying a defect based on a change amount of a characteristic frequency peak caused by a bearing defect or a shaft touch.

JP 2011-070635 A JP 2013-030015 A

However, since the measurement results of the vibration of the rolling bearing provided in the rotating machine include noise caused by the operating conditions and the installation location, the measured physical quantity to some extent even when the rolling bearing is in a normal state. fluctuate. Therefore, even if an initial defect, that is, a minute defect, occurs in the rolling bearing, it is difficult to distinguish it from the normal state and cannot be sufficiently detected.

The present invention has been made in view of the above problems, and includes a rolling bearing defect detection system, a rolling bearing defect model creation program, and a rolling bearing defect detection program that accurately detect the occurrence of minute defects in the rolling bearing. The purpose is to provide.

In order to achieve the above object, a defect detection system according to one aspect of the present invention includes a model learning unit that learns a model using a rolling bearing whose defect state is known, and a defect in the rolling bearing using the learned model. It is a defect detection system of a rolling bearing provided with the detection part to detect.

The model learning unit acquires a normal signal that is a measurement result in a state in which a shaft supported by a normal bearing that is a defect-free rolling bearing is rotated, and is the same type of rolling as the normal bearing having a relatively small defect. A first signal that is a measurement result in a state where the shaft body supported by the first bearing that is a bearing is rotated is acquired, and is a rolling bearing of the same type as the normal bearing and has a larger defect than the first bearing. A reference information acquisition unit that acquires a second signal that is a measurement result in a state in which a shaft supported by two bearings is rotated, and information based on the normal signal, the first signal, and the second signal A division unit that divides a plurality of segments by time length, and an initial value that includes an initial number of initial feature values by calculating a feature value for each segment of the normal signal, the first signal, and the second signal Create feature vector Classifying the normal signal and the first signal, and the normal signal and the second signal by supervised learning using the generated initial feature vector, and determining the contribution when classifying the first signal. A first feature amount that is a high feature amount, a high contribution feature amount extraction unit that extracts a second feature amount that is a feature amount having a high contribution in classifying the second signal from an initial feature amount, and the first feature A first normal vector based on the normal signal using a quantity, a first feature vector based on the first signal, and a second normal vector based on the normal signal using the second feature quantity, A vector reconstruction unit for reconstructing a second feature vector based on the second signal, and learning the first normal model and the second normal model using the first normal vector and the second normal vector, respectively. A normal model learning unit configured to calculate a first defect degree for the first normal model using the first feature vector, and a second defect degree for the second normal model using the second feature vector A reference defect degree calculation unit, an overall normal model creation unit that creates an overall normal model using the first defect degree and the second defect degree as input vectors, and the first defect degree and the second defect degree are input. As a vector, a normal signal defect degree calculation unit that calculates a normal signal defect degree that is a defect degree of a normal signal from the overall normal model, and a defect detection threshold setting unit that sets a defect detection threshold based on the normal signal defect degree Prepare.

The detection unit includes an evaluation information acquisition unit that acquires an evaluation signal that is a measurement result of a state in which a shaft body supported by an evaluation bearing that is a rolling bearing of the same type as the normal bearing is rotated; An evaluation dividing unit that divides into a plurality of segments by time length, and a first evaluation vector using the first feature amount and a second feature amount for each segment of the divided evaluation signal An evaluation vector generation unit for generating two evaluation vectors, a first evaluation defect degree for the first normal model of the evaluation signal using the first evaluation vector, and the evaluation signal using the second evaluation vector An evaluation defect degree calculation unit for calculating a second evaluation defect degree with respect to the second normal model, the first evaluation defect degree, and the second evaluation defect degree as input vectors for evaluation from the general normal model The evaluation signal defect degree calculation unit that calculates the defect degree of the evaluation signal is compared with the defect detection threshold, and the defect degree is detected as a defect out of all the information included in the evaluation signal. A defect rate calculating unit that calculates a ratio exceeding the threshold as a defect rate; and a determination unit that determines that the rolling bearing that has obtained the evaluation signal has a defect when the defect rate exceeds a predetermined defect rate threshold.

In order to achieve the above object, a defect model creation program that is one of the present invention is a normal signal that is a measurement result in a state in which a shaft supported by a normal bearing that is a rolling bearing without a defect is rotated. To obtain a first signal that is a measurement result of a state in which a shaft body supported by a first bearing that is a rolling bearing of the same type as the normal bearing having a relatively small defect is rotated, and the normal bearing A reference information acquisition unit that acquires a second signal that is a measurement result of a state in which a shaft body supported by a second bearing that is a rolling bearing of the same type and has a larger defect than the first bearing is rotated, and the normal signal A division unit that divides information based on the first signal and the second signal into a plurality of segments with a predetermined time length; and for each segment of the normal signal, the first signal, and the second signal Feature quantity An initial feature vector creation unit that creates an initial feature vector composed of an initial number of initial feature quantities by calculation, and uses the created initial feature vector to teach the normal signal and the first signal, and the normal signal and the second signal, respectively. The initial feature value is the first feature value, which is a feature value with a high contribution when classifying the first signal, and the second feature value with a high contribution value when the second signal is classified. A high-contribution feature amount extraction unit that respectively extracts the first feature vector, the first normal vector based on the normal signal using the first feature amount, and the first feature vector based on the first signal, and the second feature A second normal vector based on the normal signal using a quantity, a vector reconstruction unit for reconstructing a second feature vector based on the second signal, the first normal vector, and the second normal vector. A normal model learning unit that uses the first feature vector to create a first normal model and a second normal model by learning; a first defect degree for the first normal model using the first feature vector; and the second feature vector. A reference defect degree calculation unit that calculates a second defect degree with respect to the second normal model using, and an overall normal model creation part that creates an overall normal model using the first defect degree and the second defect degree as input vectors A normal signal defect degree calculation unit that calculates a normal signal defect degree that is a defect degree of a normal signal from the overall normal model using the first defect degree and the second defect degree as input vectors; and the normal signal defect And a defect detection threshold value setting unit that sets a defect detection threshold value based on the degree, and each processing unit is realized by causing the computer to execute each process.

In order to achieve the above object, a defect detection program according to one aspect of the present invention is a defect detection program for detecting a defect in a rolling bearing using a defect model created by the defect model creation program. An evaluation information acquisition unit that acquires an evaluation signal that is a measurement result in a state in which a shaft body supported by an evaluation bearing that is a rolling bearing to be measured is rotated; and a plurality of segments of the evaluation signal for a predetermined time length An evaluation dividing unit that divides the evaluation signal into two segments, and an evaluation that generates a first evaluation vector using the first feature amount and a second evaluation vector using the second feature amount for each segment of the divided evaluation signal A vector generation unit; a first evaluation defect degree of the evaluation signal with respect to the first normal model using the first evaluation vector; and the evaluation signal using the second evaluation vector. An evaluation defect degree calculation unit for calculating a second evaluation defect degree with respect to the second normal model, a defect of an evaluation signal from the general normal model, with the first evaluation defect degree and the second evaluation defect degree as input vectors An evaluation signal defect degree calculation unit that calculates an evaluation signal defect degree that is a degree, and the evaluation signal defect degree is compared with the defect detection threshold value, and the defect degree exceeds the defect detection threshold value among all information included in the evaluation signal A defect rate calculation unit that calculates a ratio as a defect rate, and a determination unit that determines that the rolling bearing that has obtained the evaluation signal has a defect when the defect rate exceeds a predetermined defect rate threshold, and performs each process. The above processing units are realized by causing the computer to execute them.

According to these, it becomes possible to improve the defect detection accuracy of the rolling bearing. Note that the execution of the recording medium in which the program is recorded also corresponds to the implementation of the present invention.

According to the present invention, it is possible to improve the defect detection accuracy, for example, by making it possible to detect defects that are lighter than those of the prior art.

It is a figure which shows the outline of the measuring apparatus which concerns on this Embodiment. It is a block diagram which shows the function structure of the defect detection system which concerns on this Embodiment. It is a figure which shows typically the division | segmentation state of the division part which concerns on this Embodiment. It is a figure which shows the case where a model learning part and a detection part become a different body.

Next, embodiments of a rolling bearing defect detection system, a rolling bearing defect model creation program, and a rolling bearing defect detection program according to the present invention will be described with reference to the drawings. It should be noted that each of the embodiments described below shows a comprehensive or specific example. The numerical values, shapes, materials, constituent elements, arrangement positions and connecting forms of the constituent elements, steps, order of steps, and the like shown in the following embodiments are merely examples, and are not intended to limit the present invention. In addition, among the constituent elements in the following embodiments, constituent elements that are not described in the independent claims indicating the highest concept are described as optional constituent elements.

Also, the drawings are schematic diagrams in which emphasis, omission, and ratio adjustment are performed as appropriate to show the present invention, and may differ from actual shapes, positional relationships, and ratios.

FIG. 1 is a diagram showing an outline of a measuring device, in which a diagram shown in part (a) is a front view of the measuring device, and a diagram shown in part (b) is a sectional side view corresponding to the part (a).

As shown in the figure, a measuring apparatus 200 includes a base body 205, a shaft body 201, a target bearing 210 that is a rolling bearing to be measured, a driving device 202, a weight 203, a biasing member 204, and a measurement sensor. (Described later) and a recording device 206. In the case of the present embodiment, the measuring apparatus 200 employs a structure in which the shaft body 201 is supported by the target bearing 210 and the auxiliary bearing 211 that is guaranteed not to be defective.

The base body 205 is a member that is a structural basis of the measuring apparatus 200, and is a member that rotatably holds the shaft body 201.

The shaft body 201 is a rod-like member that is rotatably held by the base body 205 via the target bearing 210 or the like. The shaft body 201 is not particularly limited in material and length as long as it has a shape that fits the inner ring of the target bearing 210, but it is preferable to match the actual usage of the target bearing 210 as much as possible.

The target bearing 210 is a rolling bearing to be measured. The target bearing 210 is not particularly limited as long as it is a rolling bearing, but in the case of the present embodiment, the target bearing 210 is an angular ball bearing. Further, as the target bearing 210, a normal bearing which is a rolling bearing having no defect, a first bearing which is a rolling bearing having a relatively small defect, and a second bearing having a defect larger than the first bearing are prepared. In the measuring apparatus 200, the same type of rolling bearing can be used as the target bearing 210, and the rolling bearings in different defective states can be exchanged and used.

Normal bearings are unused rolling bearings that are recognized as having no scratches by product inspection. When the short axis radius of the elliptical contact portion generated on the ball of the target bearing 210 and the outer ring raceway surface is defined as b when the load is applied, the first bearing has, for example, a cylindrical hole with a diameter of 2b that is twice that of the inner circumference of the outer ring. It is a bearing provided on the surface. The second bearing is, for example, a bearing in which a cylindrical hole having a diameter that can be detected by a method of specifying a defect by a change amount of a characteristic frequency peak, which is a conventional defect detection method, is provided on the inner periphery of the outer ring. In the present embodiment, the diameter of the cylindrical hole provided in the second bearing is 6b. The holes were formed by electric discharge machining.

Here, “defects” used in the present specification and claims include holes, grooves, scratches and the like that are artificially provided as described above, and are further generated by using a bearing. The term includes fatigue peeling and fatigue damage, lubricant deterioration, and surface roughness due to seal failure.

In the case of the present embodiment, the auxiliary bearing 211 is provided to stably rotate the shaft body 201, and a bearing in the same state as a normal bearing is employed.

The drive device 202 is a device that drives the shaft body 201 to rotate. The type of the driving device 202 is not particularly limited, but in the case of the present embodiment, the drive device 202 is separated from the base body 205 so that the vibration of the motor as a power source is not transmitted to the shaft body 201 as much as possible. The driving device 202 receives power from a fixed motor (not shown) via a belt. The driving device 202 is connected to the shaft body 201 by a joint provided with a vibration isolating member.

The weight 203 is a member that applies a vertically downward load to the shaft body 201 in one of the radial directions. The weight of the weight 203 is preferably close to the usage of the rolling bearing to be evaluated.

The urging member 204 is a member that applies a load to the shaft body 201 in the axial direction. The type of the urging member 204 is not particularly limited, but in the case of the present embodiment, a leaf spring is employed. The load applied to the shaft body 201 by the urging member 204 is the same as that of the weight 203, and is preferably close to the usage mode of the rolling bearing to be evaluated.

The measurement sensor is not particularly limited and may be a sensor that measures sound, but in the case of the present embodiment, the measurement sensor is vibration (vibration acceleration) caused by rotation of the shaft body 201 due to the target bearing 210. ). However, since it is generally considered difficult to attach the measurement sensor directly to the target bearing 210, in the case of the present embodiment, the vibration sensor is attached to the base body 205, and only the vibration of the target bearing 210 is detected. Also measure vibrations caused by other factors.

The type of the measurement sensor is not particularly limited, and it is preferable to employ the same type of sensor as that used to detect a defect in the rolling bearing. In the case of the present embodiment, a sensor that measures vibration acceleration in a uniaxial direction is employed as the measurement sensor. The number and location of the measurement sensors attached to the measuring apparatus 200 are not limited, but in the case of the present embodiment, the radial direction is the axial direction of the shaft body 201 and is added to the shaft body 201 by the weight 203. A first sensor 221 that measures vibration in the direction of the load (Z direction in FIG. 1), and a direction that is in the radial direction and perpendicular to the direction of vibration that is measured by the first sensor 221 and that is in the horizontal plane and outside the device A second sensor 222 that measures vibration in a direction in which the restraining force from the shaft is weak, and a third sensor 223 that acquires vibration in the axial direction of the shaft body 201.

As described above, it is possible to improve the detection accuracy of the defect of the rolling bearing by measuring the vibration in the three-dimensional direction about the target bearing 210. In particular, by measuring the vibration in a direction that is in the horizontal plane in the radial direction with respect to the shaft body 201 and in which the restraining force from the outside of the apparatus is weak, it is possible to improve the detection accuracy of minute defects that cannot be detected by the conventional method. .

The recording device 206 is a device that records a signal from the measurement sensor. In this embodiment, signals from the first sensor 221, the second sensor 222, and the third sensor 223 are individually recorded. As a specific recording method, an analog signal from a measurement sensor is digitized at a sampling frequency of 50 kHz, and the signal is recorded for a time of about 5 to 60 seconds (20 seconds in this embodiment). In addition, every time a signal was acquired for the normal bearing, the first bearing, and the second bearing, the bearings were recombined, and the signal was recorded 11 times for each. Thereby, the influence which the recombination of the object bearing 210 has on a signal can be suppressed, and the detection accuracy of a defect can be improved. The rotational speed of the shaft body 201 is not particularly limited, and is limited by the type of the target bearing 210. In the present embodiment, the rotational speed selected from a range of 1000 rpm to 2000 rpm. It was rotated with.

Next, the configuration of the defect detection system will be described. FIG. 2 is a block diagram illustrating a functional configuration of the defect detection system.

The defect detection system 100 uses a model learning unit 101 that learns a model using a signal of a rolling bearing whose defect state is known measured using the above-described measuring apparatus 200 and the like, and a model learned by the model learning unit 101. A model learning unit 101 includes a reference information acquisition unit 102, a division unit 103, and an initial feature vector generation unit. The detection unit 151 detects the presence or absence of a defect from a rolling bearing signal whose defect state is unknown. 104, a high contribution feature amount extraction unit 105, a vector reconstruction unit 106, a normal model learning unit 107, a reference defect degree calculation unit 108, an overall normal model creation unit 109, and a normal signal defect degree calculation unit 110 A defect detection threshold value setting unit 111, and the detection unit 151 includes an evaluation information acquisition unit 152, an evaluation division unit 153, and an evaluation vector generation unit 15. If, evaluation defect calculation unit 155, an evaluation signal defect calculation unit 156, and a defect rate calculation unit 157, a determining unit 158.

The reference information acquisition unit 102 is a processing unit that acquires a signal from the recording device 206 of the measurement device 200. One of the signals is a normal signal as a signal when the target bearing 210 is a normal bearing without a defect. The other one of the signals is the first signal as a signal when the target bearing 210 is a first bearing that is a rolling bearing of the same type as a normal bearing in which a small-diameter hole that is a relatively small defect is artificially provided. One signal. The other one of the signals is when the target bearing 210 is a rolling bearing of the same kind as a normal bearing and a second bearing in which a rectangular hole that is a larger defect than the first bearing is artificially provided. It is the 2nd signal as a signal.

FIG. 3 is a diagram schematically showing the division state of the division unit.

As shown in the figure, the dividing unit 103 divides the normal signal, the first signal, and the second signal acquired by the reference information acquisition unit 102 by a predetermined time length, respectively, so that a normal segment group and a first segment group are obtained. And a processing unit for generating a second segment group. A vertical line overlapping the character in the figure indicates that it is after division. The predetermined time length for dividing the signal by the dividing unit 103 is not particularly limited, but is set to be at least a time length enough for the shaft body 201 to rotate a plurality of times. In the case of the present embodiment, each signal is divided into a plurality of segments with a time length corresponding to five rotations of the shaft body 201.

In the present embodiment, the dividing unit 103 extracts a plurality of frequency bands by applying a band pass filter to the normal segment group, the first segment group, and the second segment group, and generates new information. . The frequency bandwidth and the number of extractions extracted by the dividing unit 103 are not particularly limited, but in the case of the present embodiment, 20 Hz to 200 Hz (frequency band 01), 1000 Hz to 5000 Hz (frequency band 02), and 5000 Hz. Three types of frequency of 20000 Hz or less (frequency band 03) are extracted. Since the initial feature vector is also created for raw data to which the bandpass filter is not applied, four frequency bandwidth segments are created in the present embodiment.

In the case of the present embodiment, the dividing unit 103 performs Fourier transform by fast Fourier transform on the normal segment group, the first segment group, and the second segment group that are the time domain information including the extracted new information. Frequency domain information is generated and newly included in the normal segment group, the first segment group, and the second segment group. Further, the dividing unit 103 performs Fourier transform on the frequency domain information of each segment by fast Fourier transform to generate quefrency region information, and newly includes it in the normal segment group, the first segment group, and the second segment group.

As described above, the normal segment group includes the segment divided for the four types of frequency bands, the frequency domain segment that has undergone Fourier transform, and the segment of the quefrency domain that has undergone Fourier transform, as a result of the arithmetic processing of the division unit 103. Twelve types of segments are included. Moreover, the normal signal which becomes the origin of a normal segment group contains the signal obtained from the 1st sensor 221, the 2nd sensor 222, and the 3rd sensor 223, and the division | segmentation part 103 processes the information from each sensor separately. Therefore, the normal segment group includes 36 types of segments. The same applies to the first segment group and the second segment group.

In addition, the normal signal, the first signal, and the second signal include the signal from the first sensor 221, the signal from the second sensor 222, and the signal from the third sensor 223, respectively. The dividing unit 103 performs the above division.

The initial feature vector creation unit 104 calculates a feature amount for each segment of the normal segment group, the first segment group, and the second segment group generated by the dividing unit 103, and an initial number of initial feature amounts based on the calculation result Create an initial feature vector consisting of The type of feature quantity calculated from each segment and the number to be calculated are not particularly limited. For example, various statistics can be selected and used. In the case of the present embodiment, five types of statistics, which are an effective value, a maximum value, a crest factor, a kurtosis, and a skewness, are selected as feature amounts calculated from each segment.

In the case of the present embodiment, the frequency domain statistic is calculated with respect to the frequency domain waveform obtained by performing Fourier transform after envelope processing of each segment. Further, the statistic of the quefrency region is calculated with respect to the waveform of the quefrency region obtained by performing Fourier transform again on the frequency region information after the envelope processing.

Since the initial feature vector creation unit 104 of the present embodiment creates the initial feature vector, the initial feature amount is information from any of the first sensor 221, the second sensor 222, and the third sensor 223. 3 types, 4 types of information in frequency band, 3 types of information in time domain information, frequency domain information, and quefrency region information, and feature quantity calculated as statistic The initial number that is the number of the initial feature amounts is 180 including five types. That is, in the case of this embodiment, the initial feature vector creation unit 104 creates an initial feature vector composed of 180 feature amounts.

It should be noted that the initial feature vector is not created for all the acquired signals, but a plurality of initial feature vectors may be selected at random for each segment of the normal signal, the first signal, and the second signal.

The high contribution feature amount extraction unit 105 classifies the normal signal and the first signal, and the normal signal and the second signal by supervised learning using the created initial feature vector, and contributes to classify the first signal. A first feature amount that is a high feature amount and a second feature amount that is a feature amount having a high contribution in classifying the second signal, respectively, from the initial feature amount.

The method for extracting feature amounts having a high contribution is not particularly limited, but the contribution of each feature is calculated for each of the first signal and the second signal by supervised machine learning, and the contribution is calculated. What is necessary is just to extract a predetermined number (for example, 10 pieces) from the upper rank of.

In the case of the present embodiment, the high contribution feature amount extraction unit 105 uses a random forest that is one of supervised machine learning, extracts a first feature amount based on the normal signal and the first signal, A second feature amount is extracted based on the signal and the second signal.

Random forest is a classification method using decision trees. A plurality of training sets are created by restoration extraction from the input data, and each training set is classified by a decision tree. The final classification is determined by majority vote for the classification result by the decision tree for each training set.

The vector reconstruction unit 106 reconstructs the first normal vector for each normal segment group based on the normal signal using the first feature amount extracted by the high-contribution feature amount extraction unit 105, and performs each first based on the first signal. The first feature vector is reconstructed for the segment group, the second feature vector is used to reconstruct the second normal vector for each normal segment group based on the normal signal, and the second feature vector is second for each second segment group based on the second signal. It is a processing unit for reconstructing a feature vector.

In the case of this embodiment, the high contribution feature amount extraction unit 105 extracts the top ten first feature amounts and second feature amounts having a high contribution degree from 378 feature amounts. Accordingly, the vector reconstruction unit 106 reconstructs the first normal vector and the first feature vector based on the ten first feature amounts, and constructs the second normal vector and the second feature vector based on the second feature amount. Yes.

The normal model learning unit 107 creates a first normal model and a second normal model by learning using the first normal vector and the second normal vector created by the vector reconstruction unit 106, respectively.

In the case of the present embodiment, the normal model learning unit 107 generates a function for calculating the reference defect degree by fitting an unsupervised machine learning algorithm to the first normal vector and the second normal vector. ing. Specifically, a local authorial factor (hereinafter referred to as LOF) is employed.

The reference defect degree calculation unit 108 is a processing unit that calculates the first defect degree for the first normal model using the first feature vector and the second defect degree for the second normal model using the second feature vector. .

The comprehensive normal model creation unit 109 is a processing unit that creates a comprehensive normal model using the first defect degree and the second defect degree calculated by the reference defect degree calculation unit 108 as input vectors.

In the case of the present embodiment, the general normal model creation unit 109 adopts the LOF similarly to the normal model learning unit 107.

The normal signal defect degree calculation unit 110 calculates a normal signal defect degree that is a defect degree of a normal signal from the overall normal model created by the overall normal model creation unit 109 using the first defect degree and the second defect degree as input vectors. Is a processing unit.

The defect detection threshold value setting unit 111 is a processing unit that sets a defect detection threshold value based on the normal signal defect degree. Specifically, for example, the defect detection threshold setting unit 111 sets an average of all normal signal defect degrees + 5 × standard deviation as the defect detection threshold.

The defect detection threshold setting unit 111 may set the defect detection threshold based on information from the outside, for example, information input by a person.

The evaluation information acquisition unit 152 is a processing unit that acquires an evaluation signal that is a measurement result of a state in which a shaft body supported by an evaluation bearing that is a rolling bearing of the same type as the normal bearing used in the model learning unit 101 is rotated. is there.

In the case of the present embodiment, a measurement sensor of the same type as the measurement sensor used in the model learning unit 101 is used in an actual machine 300 such as an industrial machine in which the same type of rolling bearing as that of a normal bearing is used, and the same number as that of the measuring apparatus 200 is used. And a plurality of measurement sensors mounted in an arrangement are mounted in the vicinity of the evaluation bearing. Then, the evaluation information acquisition unit 152 acquires the signal recorded in the real machine recording device 306 as in the measurement device 200.

The evaluation dividing unit 153 divides the evaluation signal acquired by the evaluation information acquiring unit 152 into a plurality of segments with a predetermined time length that is the same as or substantially the same as that of the model learning unit 101. In the case of the present embodiment, the evaluation dividing unit 153 extracts frequency band information in the same or substantially the same frequency band as the dividing unit 103. Further, the evaluation division unit 153 performs frequency transformation information on each segment to create frequency domain information, and further performs Fourier transformation to create quefrency region information similarly to the division unit 103.

Note that if the feature amount is not included in the first feature amount or the second feature amount extracted by the high contribution feature amount extraction unit 105, the evaluation division unit 153 may not create the corresponding information. Absent. Specifically, for example, when the first feature value or the second feature value does not include a feature value that is quefrency region information, the evaluation division unit 153 does not need to perform Fourier transform on the frequency region information. As a result, the processing time of the evaluation division unit 153 can be shortened.

The evaluation vector generation unit 154 uses the first evaluation vector and the second feature amount for each segment of the evaluation signal divided by the evaluation division unit 153, using the first feature amount obtained from the model learning unit 101. It is a process part which produces | generates a 2nd evaluation vector. Since the first feature amount and the second feature amount have a smaller number of feature amounts than the initial feature amount, the evaluation vector generation unit 154 can generate each evaluation vector in a short time.

The evaluation defect degree calculation unit 155 acquires the first normal model and the second normal model from the model learning unit 101, and uses the first evaluation vector created by the evaluation vector generation unit 154 to evaluate the first normal model of the evaluation signal. And a second evaluation defect degree for the second normal model of the evaluation signal using the second evaluation vector.

The evaluation signal defect degree calculation unit 156 uses the first evaluation defect degree and the second evaluation defect degree calculated by the evaluation defect degree calculation unit 155 as input vectors, and the defect of the evaluation signal from the general normal model acquired from the model learning unit 101 It is a processing unit for calculating an evaluation signal defect degree which is a degree.

The defect rate calculation unit 157 compares the evaluation signal defect degree calculated by the evaluation signal defect degree calculation unit 156 with the defect detection threshold acquired from the model learning unit 101, and the defect degree is defective among all the information included in the evaluation signal. It is a processing unit that calculates a ratio exceeding the detection threshold as a defect rate.

The determination unit 158 is a processing unit that determines that there is a defect in the evaluation bearing that is the rolling bearing that has acquired the evaluation signal when the defect rate calculated by the defect rate calculation unit 157 exceeds a predetermined defect rate threshold.

In the case of the present embodiment, the detection unit 151 includes a notification unit 159. When the determination unit 158 determines that the evaluation bearing is defective, the notification unit 159 is a processing unit that notifies the information.

The method of notification is not particularly limited, but it may be notified that there is a defect by transmitting information to sound, light, image, video, and other computers.

According to the above-described embodiment, the defect detection accuracy is higher than that of the prior art even in adverse conditions such as a situation in which the operation status of the actual machine including the evaluation bearing varies and a situation in which vibration from other machines is transmitted. In addition, it is possible to detect defects that are lighter than those of the prior art. Therefore, it is possible to reduce the influence on the surrounding structure due to oversight of minute defects generated in the rolling bearing in the actual machine. According to this, since it is possible to avoid the repair of the peripheral structure of the rolling bearing, it is possible to improve the operating rate of the actual machine.

The present invention is not limited to the above embodiment. For example, another embodiment realized by arbitrarily combining the components described in this specification and excluding some of the components may be used as an embodiment of the present invention. In addition, the present invention includes modifications obtained by making various modifications conceivable by those skilled in the art without departing from the gist of the present invention, that is, the meaning of the words described in the claims. It is.

For example, in the present embodiment, the defect detection system 100 including the model learning unit 101 and the detection unit 151 has been described. However, as illustrated in FIG. 4, the model learning unit 101 and the detection unit 151 are separate. It doesn't matter. Specifically, a defect model creation program 301 that causes the measurement computer 310 to execute each processing unit included in the model learning unit 101 and a defect detection program 312 that causes the evaluation computer 311 to execute each processing unit included in the detection unit 151 are used. Can be realized.

In this case, the measurement computer 310 and the evaluation computer 311 may directly exchange necessary information, and the measurement computer 310 may send a plurality of information to the server 313 as shown in FIG. The information may be downloaded from the server 313 and the evaluation computer 311 may download information suitable for the rolling bearing to be evaluated.

In addition, the signals used by the model learning unit 101 are not limited to the normal signal, the first signal, and the second signal, and the third signal, the fourth signal, and the like that are different from each other in the defect state of the rolling bearing. You can use it.

Also, the signal measured by the measurement sensor is not limited to vibration acceleration, and may be vibration displacement, vibration speed, or the like. Further, not only vibration but also sound may be measured and used as a signal.

Also, the frequency band to be extracted may be arbitrarily selected, and the extracted frequency bands may overlap.

Further, it is not always necessary to execute the Fourier transform, and the frequency domain information may be used only for creating the quefrency domain information, and the feature quantity may not be included in the frequency domain.

In FIG. 2, each processing unit is described separately. However, if there is a module common to each processing unit, they may be shared.

The present invention is applicable to various machines equipped with rolling bearings.

DESCRIPTION OF SYMBOLS 100 Defect detection system 101 Model learning part 102 Reference | standard information acquisition part 103 Division | segmentation part 104 Initial feature vector preparation part 105 High contribution feature-value extraction part 106 Vector reconstruction part 107 Normal model learning part 108 Reference | standard defect degree calculation part 109 Comprehensive normal model preparation Unit 110 normal signal defect degree calculation unit 111 defect detection threshold setting unit 151 detection unit 152 evaluation information acquisition unit 153 evaluation division unit 154 evaluation vector generation unit 155 evaluation defect degree calculation unit 156 evaluation signal defect degree calculation unit 157 defect rate calculation unit 158 Determination unit 159 Notifying unit 200 Measuring device 201 Shaft body 202 Drive device 203 Weight 204 Biasing member 205 Base 206 Recording device 210 Target bearing 211 Auxiliary bearing 221 First sensor 222 Second sensor 223 Third sensor 300 Actual machine 301 Defect model creation program 306 Actual machine record Location 310 measuring computer 311 evaluation computer 312 defect detection program 313 server

Claims

A rolling bearing defect detection system comprising a model learning unit that learns a model using a rolling bearing whose defect state is known, and a detection unit that detects a defect of the rolling bearing using the learned model,
The model learning unit
A first bearing which is a rolling bearing of the same type as the normal bearing having a relatively small defect, which obtains a normal signal as a measurement result in a state in which the shaft body supported by the normal bearing which is a rolling bearing having no defect is rotated. The first signal, which is the measurement result of the state where the shaft body supported by the shaft is rotated, is acquired and is supported by a second bearing that is a rolling bearing of the same type as the normal bearing and has a larger defect than the first bearing. A reference information acquisition unit for acquiring a second signal that is a measurement result in a state where the shaft body is rotated;
A division unit that divides information based on the normal signal, the first signal, and the second signal into a plurality of segments with a predetermined time length;
An initial feature vector creating unit that creates a feature quantity for each segment of the normal signal, the first signal, and the second signal and creates an initial feature vector including an initial number of initial feature quantities;
Using the created initial feature vector, the normal signal and the first signal, and the normal signal and the second signal are classified by supervised learning, respectively. A high-contribution feature quantity extraction unit that extracts a second feature quantity, which is a feature quantity having a high contribution in classifying the quantity and the second signal, from the initial feature quantity;
Reconstructing a first normal vector based on the normal signal using the first feature quantity and a first feature vector based on the first signal, and a second normal vector based on the normal signal using the second feature quantity A vector and a vector reconstruction unit for reconstructing a second feature vector based on the second signal;
A normal model learning unit that creates a first normal model and a second normal model by learning using the first normal vector and the second normal vector, respectively;
A reference defect degree calculation unit that calculates a first defect degree for the first normal model using the first feature vector and a second defect degree for the second normal model using the second feature vector;
An overall normal model creating unit that creates an overall normal model using the first defect degree and the second defect degree as input vectors;
A normal signal defect degree calculation unit that calculates a normal signal defect degree that is a defect degree of a normal signal from the overall normal model, using the first defect degree and the second defect degree as input vectors;
A defect detection threshold setting unit that sets a defect detection threshold based on the normal signal defect degree,
The detector is
An evaluation information acquisition unit that acquires an evaluation signal that is a measurement result in a state in which a shaft body supported by an evaluation bearing that is a rolling bearing of the same type as the normal bearing is rotated;
An evaluation dividing unit that divides the evaluation signal into a plurality of segments with a predetermined time length;
An evaluation vector generation unit that generates a first evaluation vector using the first feature amount and a second evaluation vector using the second feature amount for each segment of the divided evaluation signal;
A first evaluation defect degree for the first normal model of the evaluation signal using the first evaluation vector, and a second evaluation defect degree for the second normal model of the evaluation signal using the second evaluation vector. An evaluation defect degree calculation unit to calculate,
An evaluation signal defect degree calculation unit that calculates an evaluation signal defect degree that is a defect degree of an evaluation signal from the overall normal model, using the first evaluation defect degree and the second evaluation defect degree as input vectors;
A defect rate calculating unit that compares the evaluation signal defect degree with the defect detection threshold value, and calculates a ratio of the defect degree exceeding the defect detection threshold value as a defect rate among all information included in the evaluation signal;
A defect detection system comprising: a determination unit that determines that the rolling bearing that has obtained the evaluation signal is defective when the defect rate exceeds a predetermined defect rate threshold.
The defect detection system according to claim 1, wherein each of the normal signal, the first signal, and the second signal is vibration information.
The normal signal, the first signal, the second signal, and the evaluation signal are axial vibration information that is vibration information in the axial direction of the rolling bearing to be measured, and vibration information in the first radial direction, respectively. The defect detection system according to claim 2, comprising vibration information and second vibration information that is vibration information in a second radial direction that intersects the axial direction and the first radial direction.
The second radial direction, which is the vibration direction of the second vibration, is included in a horizontal plane, and an external force is applied between the shaft body to be measured and the rolling bearing in the second radial direction. The defect detection system according to claim 3.
The dividing unit performs a Fourier transform on the normal signal, the first signal, and the second signal to generate frequency domain information,
The defect detection system according to any one of claims 1 to 4, wherein the initial feature vector creation unit creates an initial feature vector as one of the feature information that is the frequency domain information.
The dividing unit creates quefrency region information obtained by performing further Fourier transform on the frequency region,
The defect detection system according to claim 5, wherein the initial feature vector creation unit creates an initial feature vector as one of the initial feature amounts, which is the kerf region information.
The defect detection system according to claim 1, wherein the parameter of the initial feature amount includes at least one of an effective value, a maximum value, a crest factor, a modulation value, a kurtosis, and a skewness. .
The dividing unit extracts information of a plurality of frequency bands by applying a band pass filter to the normal signal, the first signal, and the second signal,
The defect detection system according to any one of claims 1 to 7, wherein the evaluation division unit applies a similar bandpass filter to the evaluation signal to extract information on a plurality of the frequency bands.
A first bearing which is a rolling bearing of the same type as the normal bearing having a relatively small defect, which obtains a normal signal as a measurement result in a state in which the shaft body supported by the normal bearing which is a rolling bearing having no defect is rotated. The first signal, which is the measurement result of the state where the shaft body supported by the shaft is rotated, is acquired and is supported by a second bearing that is a rolling bearing of the same type as the normal bearing and has a larger defect than the first bearing. A reference information acquisition unit for acquiring a second signal that is a measurement result in a state where the shaft body is rotated;
A division unit that divides information based on the normal signal, the first signal, and the second signal into a plurality of segments with a predetermined time length;
An initial feature vector creating unit that creates a feature quantity for each segment of the normal signal, the first signal, and the second signal and creates an initial feature vector including an initial number of initial feature quantities;
Using the created initial feature vector, the normal signal and the first signal, and the normal signal and the second signal are classified by supervised learning, respectively. A high-contribution feature quantity extraction unit that extracts a second feature quantity, which is a feature quantity having a high contribution in classifying the quantity and the second signal, from the initial feature quantity;
Reconstructing a first normal vector based on the normal signal using the first feature quantity and a first feature vector based on the first signal, and a second normal vector based on the normal signal using the second feature quantity A vector and a vector reconstruction unit for reconstructing a second feature vector based on the second signal;
A normal model learning unit that creates a first normal model and a second normal model by learning using the first normal vector and the second normal vector, respectively;
A reference defect degree calculation unit that calculates a first defect degree for the first normal model using the first feature vector and a second defect degree for the second normal model using the second feature vector;
An overall normal model creating unit that creates an overall normal model using the first defect degree and the second defect degree as input vectors;
A normal signal defect degree calculation unit that calculates a normal signal defect degree that is a defect degree of a normal signal from the overall normal model, using the first defect degree and the second defect degree as input vectors;
A defect detection threshold setting unit that sets a defect detection threshold based on the normal signal defect degree,
A defect model creation program for realizing each processing unit by causing a computer to execute each processing.
A defect detection program for detecting a defect of a rolling bearing using a defect model created by the defect model creation program according to claim 9,
An evaluation information acquisition unit that acquires an evaluation signal that is a measurement result in a state in which a shaft body supported by an evaluation bearing that is a rolling bearing to be measured is rotated;
An evaluation dividing unit that divides the evaluation signal into a plurality of segments with a predetermined time length;
An evaluation vector generation unit that generates a first evaluation vector using the first feature amount and a second evaluation vector using the second feature amount for each segment of the divided evaluation signal;
A first evaluation defect degree for the first normal model of the evaluation signal using the first evaluation vector, and a second evaluation defect degree for the second normal model of the evaluation signal using the second evaluation vector. An evaluation defect degree calculation unit to calculate,
An evaluation signal defect degree calculation unit that calculates an evaluation signal defect degree that is a defect degree of an evaluation signal from the overall normal model, using the first evaluation defect degree and the second evaluation defect degree as input vectors;
A defect rate calculating unit that compares the evaluation signal defect degree with the defect detection threshold value, and calculates a ratio of the defect degree exceeding the defect detection threshold value as a defect rate among all information included in the evaluation signal;
A determination unit that determines that the rolling bearing that acquired the evaluation signal has a defect when the defect rate exceeds a predetermined defect rate threshold,
A rolling bearing defect detection program that realizes each processing unit by causing a computer to execute each processing.