WO2021117752A1

WO2021117752A1 - Rolling bearing state monitoring method and rolling bearing state monitoring device

Info

Publication number: WO2021117752A1
Application number: PCT/JP2020/045772
Authority: WO
Inventors: 正嗣北井
Original assignee: Ntn株式会社
Priority date: 2019-12-11
Filing date: 2020-12-09
Publication date: 2021-06-17

Abstract

The present invention includes: a step for generating feature amount vectors of a first rolling bearing, which are formed of feature amounts of vibration data on the first rolling bearing; a step for performing machine-learning for the feature amount vectors of the first rolling bearing and damage sizes of the first rolling bearing to generate a first regression model for estimating the damage sizes from the feature amounts; a step for generating a feature amount vector of a second rolling bearing, which is formed of a feature amount of vibration data on the second rolling bearing; and a step for estimating, by using the first regression model, a damage size of the second rolling bearing from the feature amount vector of the second rolling bearing.

Description

Rolling bearing condition monitoring method and rolling bearing condition monitoring device

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

Bearings (rolling bearings) that are not easy to replace and are used under relatively low speed conditions, such as spindle bearings for wind power generators, are often used continuously even if initial damage occurs. On the other hand, if the size of the damaged portion increases and the displacement between the inner and outer rings of the bearing increases and the displacement exceeds the permissible range, the device itself including the rolling bearing may be destroyed.

Therefore, in order to use rolling bearings to the limit of the safe range, it is necessary to accurately grasp the damage situation. In addition, it is desirable that the damage situation can be estimated from the measurement data that can be easily measured. The machine learning device described in Japanese Patent No. 6148316 (Patent Document 1) determines the relationship between a state variable including output data of a sensor that detects the state of an industrial machine or an ambient environment and a failure occurrence, that is, a failure condition. learn. As a result, the machine learning device can output the presence / absence of failure or the degree of failure in response to the input state variable.

Furthermore, in order to use rolling bearings to the limit of the safe range, it is necessary to accurately estimate the remaining operating time (remaining life) after damage occurs. Japanese Patent Application Laid-Open No. 2017-219469 (Patent Document 1) describes the remaining life of a rolling bearing from the time of damage diagnosis until the damaged part progresses to the allowable limit size based on the effective vibration value obtained by the vibration sensor. The theoretical formula to predict is described.

Japanese Patent No. 6148316 JP-A-2017-219469

However, in Patent Document 1, the degree of failure is exemplified as an index value, and a configuration for directly estimating the size of the damaged portion, which indicates the state of damage, is not shown.

Further, in Patent Document 2, since the remaining life of the actually damaged rolling bearing is not measured, the accuracy of the above theoretical formula has not been sufficiently confirmed. Therefore, there is no guarantee that the remaining life can be accurately estimated, especially when it is predicted that it is difficult to express by a simple theoretical formula such that the form of damage progress changes as the operation time increases. I am concerned.

Therefore, a main object of the present invention is to provide a condition monitoring method and a condition monitoring device capable of accurately estimating the size of a damaged part of a rolling bearing or the remaining life of a rolling bearing.

According to a certain aspect of the present invention, the method of monitoring the state of a rolling bearing is a step of generating a regression model by machine learning based on the vibration data of the first rolling bearing and a step of generating a regression model by using the regression model of the second rolling bearing. It includes a step to monitor the condition. Each of the first rolling bearing and the second rolling bearing is arranged between the rotating wheel, the stationary wheel, and the rotating wheel and the stationary wheel, and moves on the raceway surface of the stationary wheel as the rotating wheel rotates. It has a plurality of rolling elements. Assuming that the circumferential length of the damaged portion formed on the raceway surface of the stationary wheel of the rolling bearing is the damage size, the step of generating the regression model is the vibration data of the first rolling bearing and the damage of the first rolling bearing. The step of acquiring the size, the step of generating the feature amount vector of the first rolling bearing consisting of the feature amount of the vibration data of the first rolling bearing, the feature amount vector of the first rolling bearing, and the first rolling bearing. Including the step of generating a regression model for estimating the damage size from the feature vector by performing machine learning with the damage size of. The step of monitoring the state of the second rolling bearing is a step of acquiring the vibration data of the second rolling bearing and a feature quantity vector of the second rolling bearing composed of the feature quantity of the vibration data of the second rolling bearing. It includes a step of generating and a step of estimating the damage size of the second rolling bearing from the feature vector of the second rolling bearing using a regression model.

According to the present invention, it is possible to provide a rolling bearing condition monitoring method and a condition monitoring device capable of accurately estimating the size of a damaged portion of a rolling bearing or the remaining life of a rolling bearing.

It is a figure which showed together the cross-sectional view of the rolling bearing monitored by the state monitoring device by Embodiment 1 and the block diagram of the state monitoring device. It is a figure which showed the state (stage 1) immediately after the initial damage occurred on the raceway surface. It is the figure which showed the state (stage 2) in which the initial damage progresses in the axial direction with a rolling element. It is a figure which showed the state (stage 3) in which the damage progresses in the circumferential direction. It is a graph which shows the experimental result of the time-dependent change of the damage size. It is a figure for demonstrating the time-dependent change of the damage size of FIG. It is a figure for demonstrating a segment and a feature vector. It is a figure for demonstrating the generation of the regression model by machine learning, and the estimation of the damage size using the model. It is a figure for demonstrating the evaluation of the estimation accuracy of a regression model. This is an example of a flowchart for explaining the process of generating a regression model by machine learning. This is an example of a flowchart for explaining the process of estimating the damage size using the regression model. It is a figure which showed roughly the structure of the wind power generation apparatus to which the condition monitoring apparatus which concerns on Embodiment 1 is applied. It is a figure for demonstrating the generation of the regression model by machine learning, and the estimation of the remaining life using the model. It is a figure for demonstrating the evaluation of the estimation accuracy of a regression model. This is an example of a flowchart for explaining the process of generating a regression model by machine learning. This is an example of a flowchart for explaining a process of estimating the remaining life using a regression model. It is a figure which showed roughly the structure of the wind power generation apparatus to which the condition monitoring apparatus which concerns on Embodiment 2 is applied.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, the same or corresponding elements are designated by the same reference numerals, and detailed description thereof will not be repeated.

[Embodiment 1]
(Structure of rolling bearing and condition monitoring device)
FIG. 1 is a diagram showing a cross-sectional view of a rolling bearing monitored by the condition monitoring device according to the first embodiment of the present invention and a block diagram of the condition monitoring device. In the first embodiment, a case where a rolling bearing is formed by a roller bearing in which the outer ring is a rotating wheel is typically described, but the scope of application of the present invention is limited to such a bearing condition monitoring device. The rolling bearing to be monitored may have an inner ring of a rotating wheel, a ball bearing, or the like.

With reference to FIG. 1, the rolling bearing 10 includes an inner ring 12, an outer ring 16, and a plurality of rolling elements 18. The inner ring 12 is fitted onto the non-rotating shaft body 14. The outer ring 16 is provided on the outer peripheral side of the inner ring 12 and rotates integrally with a rotating body (not shown). Each of the plurality of rolling elements 18 is a cylindrical "roller", and is interposed between the inner ring 12 and the outer ring 16 while being held at equal intervals with the adjacent rolling elements by a cage (not shown).

The inner ring 12 receives a radial load from a plurality of rolling elements 18 that are passing through the load range. In this embodiment, in the inner ring 12 which is a stationary wheel, a load region is formed on the upper side in the vertical direction (on the paper surface) with respect to the center of rotation. Then, in the normal state in which the raceway surface (outer peripheral surface of the inner ring 12) 20 is not damaged, the inner ring 12 is located at the center of the load region located in the vertically upward direction of the inner ring central axis among the plurality of rolling elements 18. Receives maximum load from passing rolling elements.

Such rolling bearings 10 may continue to be used even after initial damage such as small-scale peeling if they are used under relatively low speed conditions and / or if they are not easy to replace. There are many. On the other hand, if the size of the damaged portion is increased and the displacement between the inner and outer rings of the bearing is increased and the displacement exceeds the allowable range, the device including the rolling bearing 10 may be destroyed.

Therefore, in order to use the rolling bearing 10 to the limit of the safe range, it is necessary to grasp the size of the damaged part. However, it is difficult to measure the size of the damaged part during use. Therefore, it is desirable to estimate the size of the damaged portion by measuring measurement data that reflects the size of the damaged portion, which can be easily measured.

Therefore, the condition monitoring device 100 of the rolling bearing 10 according to the first embodiment accurately estimates the size of the damaged portion of the rolling bearing 10 from the measurement data that can be easily measured. Specifically, the condition monitoring device 100 is configured to accurately estimate the circumferential length of the damaged portion from the measurement data of the vibration acceleration of the rolling bearing 10 by using a regression model by machine learning. ..

With reference to FIG. 1 again, the condition monitoring device 100 for monitoring the state of the rolling bearing 10 includes an acceleration sensor 102 and a diagnostic unit 106.

In FIG. 1, the x direction corresponds to the circumferential direction of the rolling bearing 10, and the y direction corresponds to the axial direction of the rolling bearing 10. Further, in FIG. 1, the X direction and the Y direction correspond to the measurement directions of the acceleration sensor 102.

The acceleration sensor 102 is fixedly installed in a place where the vibration acceleration of the bearing in the bearing box or the outer diameter surface of the equipment can be measured. The Y direction corresponds to the vertical direction (or radial direction) of the outer peripheral surface (raceway surface 20) of the inner ring 12 of the rolling bearing 10. The X direction corresponds to the horizontal direction (or horizontal direction) at the installation position of the acceleration sensor 102.

The acceleration sensor 102 is a sensor for detecting the acceleration of vibration of the rolling bearing 10 in the vertical direction (radial direction). Hereinafter, in the present specification, the acceleration of this vibration is also referred to as "vibration acceleration" or simply "acceleration". The acceleration sensor 102 may be configured to detect the vibration acceleration in the horizontal direction (horizontal direction).

The diagnosis unit 106 monitors and diagnoses the state of the rolling bearing 10 based on the detected value of the acceleration sensor 102. Specifically, the diagnostic unit 106 estimates the size of the damaged portion of the rolling bearing 10 based on the detected value of the acceleration sensor 102. Hereinafter, the size of the damaged portion of the rolling bearing 10 will be described.

(Explanation of damage size)
2 to 4 are views showing a stage in which the size of the damaged portion of the rolling bearing 10 progresses. In these figures, the load range of the rolling bearing 10 is enlarged and shown. Further, in these figures, as in FIG. 1, the y direction corresponds to the axial direction of the rolling bearing 10, and the x direction corresponds to the circumferential direction of the rolling bearing 10.

FIG. 2 shows an example in which the initial damage D1 occurs on the raceway surface 20 of the inner ring 12. Hereinafter, the state before the initial damage to the raceway surface 20 is referred to as “stage 0”, and the state in which the raceway surface 20 is initially damaged (FIG. 2) is referred to as “stage 1”.

In the specification of the present application, the size of the damaged portion refers to the circumferential length of the damaged portion formed on the raceway surface of the inner ring 12 (stationary wheel). The size of the damaged portion indicates the maximum value of the length of the damaged portion in the circumferential direction. Hereinafter, the size of such a damaged portion is also simply referred to as “damage size”. That is, the damage size of the initial damage D1 is sz1 in FIG.

FIG. 3 is a diagram showing a state in which the initial damage generated on the raceway surface 20 is progressing in the axial direction. Since the portion adjacent to the initial damage D1 as shown in FIG. 2 in the circumferential direction (x direction in FIG. 3) is the portion where the rolling element comes into contact with the rolling element before or after passing through the damaged portion. , The rolling element and the raceway surface 20 of the inner ring 12 come into contact with each other on the entire surface. Therefore, the stress distribution on the contact surface becomes uniform. On the other hand, the portion adjacent to the initial damage D1 in the axial direction (y direction in FIG. 3) is a portion that comes into contact with the rolling element while the rolling element is passing through the damaged portion. Stress concentration occurs in the adjacent portion, and as a result, the surface pressure increases locally. Therefore, the initial damage D1 basically progresses along the axial direction. Damage in which such initial damage spreads mainly in the axial direction is referred to as "damage D2". The damage size of damage D2 is sz2 in FIG. Further, a state in which the initial damage mainly progresses in the axial direction, that is, a state in which the damage D2 occurs (FIG. 3) is referred to as “stage 2”.

FIG. 4 is a diagram showing a state in which the damage is progressing in the circumferential direction. When the damage D2 shown in FIG. 3 expands over the entire axial direction, the damage then expands in the circumferential direction. After such damage spreads over the entire axial direction, the damage that spreads in the circumferential direction is referred to as "damage D3". The damage size of damage D3 is sz3 in FIG. Further, a state in which the damage spreads in the entire axial direction and then progresses in the circumferential direction, that is, a state in which the damage D3 occurs (FIG. 4) is referred to as “stage 3”.

(Change in damage size over time)
Next, the time course of damage size in the actual experimental results will be described.

FIG. 5 shows the experimental results of measuring the change in damage size after a slight initial damage D1 occurs on a part of the raceway surface 20 in the plurality of rolling bearings 10. The damage size of each rolling bearing 10 is determined by, for example, periodically stopping the device including a plurality of rolling bearings 10, photographing the stationary wheels of each rolling bearing 10 with a camera, and performing image processing on the photographed image. Can be done.

As experimental conditions, 33 cylindrical roller bearings having an inner diameter of 120 mm, an outer diameter of 240 mm, and a width of 40 m were used for the rolling bearing 10. Further, a radial load of 90 kN and a rotation speed of 500 times / minute were applied to each rolling bearing 10. For these rolling bearings 10, the damage size was measured at a predetermined time as described above, and the vibration acceleration data continuously detected by the acceleration sensor 102 was collected. As described above, the damage size indicates the maximum value of the length of the damaged portion in the circumferential direction.

The horizontal axis of FIG. 5 shows the operating time of the rolling bearing 10 after the initial damage occurs on the raceway surface 20, and the vertical axis shows the damage size. In FIG. 5, the data before the initial damage occurred is omitted. FIG. 5 shows the experimental results until the initial damage progresses and the rolling bearing 10 reaches the usage limit. The reference point of the usage limit is, for example, the time when the amount of displacement between the inner and outer rings of the bearing due to damage reaches the upper limit of the allowable range of the clearance in the equipment.

As shown in FIG. 5, it can be seen that the damage size hardly changes for a while after the initial damage occurs, and rapidly expands after a certain point in time. Hereinafter, this "state in which the damage size hardly changes with respect to the operating time" is also referred to as "initial damage". In addition, "a state in which the damage size increases rapidly" is also referred to as "end of damage". The inventors have confirmed that the "early stage of damage" and the "end stage of damage" correspond to the above-mentioned damage stages 2 (FIG. 3) and stage 3 (FIG. 4), respectively. Next, with reference to FIG. 6, the relationship between the two states of damage size and the stage of damage will be described.

FIG. 6 is a diagram for explaining the time course of damage size in FIG. FIG. 6 illustrates one of the plurality of measurement data of FIG. The times t1 and t2 in FIG. 6 indicate the times when the stage 2 and the stage 3 were entered, respectively. The time t3 in FIG. 6 indicates the time when the rolling bearing 10 reaches the usage limit, that is, the time when the damage size reaches the upper limit of the allowable range.

When the initial damage D1 occurs at time t1 (stage 1), stress concentration occurs in the portion adjacent to the initial damage D1 in the axial direction. Therefore, the damage spreads mainly in the axial direction (stage 2). Therefore, the damage size sz2 indicating the circumferential length of the damage D2 does not change significantly from the damage size sz1 of the damage D1. Therefore, as shown in FIG. 6, the propagation speed of the damage size in the circumferential direction is slow and slow.

Subsequently, at time t2, when the damage spreads over the entire axial direction, the damage further spreads in the circumferential direction (stage 3). In the stage 3, the rolling element 18 passing through the damage is in a no-load state in which a load is hardly received from the raceway surface. Therefore, the vibration of the rolling element 18 becomes intense, and as shown in FIG. 6, the propagation speed of the damage size in the circumferential direction becomes faster than that of the stage 2.

After that, when the damage further expands in the circumferential direction and the amount of displacement between the inner and outer rings of the bearing, which is the amount of displacement of the distance between the inner ring 12 and the outer ring 16, becomes larger, the damage of the rolling bearing 10 becomes the limit of use of the rolling bearing 10 at a certain point. (Time t3 in FIG. 6). In the present specification, such a state is also referred to as the rolling bearing 10 reaching the usage limit. If the rolling bearing 10 exceeds the usage limit, there is a problem that, for example, the device itself including the rolling bearing 10 is destroyed. Therefore, in such a state, the operation of the device is stopped, and the rolling bearing 10 is repaired or replaced.

The reference point for the usage limit of the rolling bearing 10 is, for example, the time when the amount of displacement between the inner and outer rings of the bearing due to damage becomes the upper limit of the allowable range of clearance in the equipment. Further, it may be a time point when the damage size becomes a constant multiple of the interval of the rolling elements 18. With this setting, for example, in a certain rolling bearing 10, when the damage size becomes 1.5 to 2 times the distance between the rolling elements, the two rolling elements 18 passing through the damaged part receive almost no load. As a result, it is possible to detect the time when the vibration of the rolling bearing 10 becomes too large.

As described above, in the rolling bearing 10, even under uniform operating conditions, the form and speed of damage progression change as the operating time increases. In particular, in the final stage of damage where the rate of damage progress increases, the fluctuation of measurement data such as vibration acceleration becomes large. As described above, the relationship between the damage progression and the measured data is complicated, and it is difficult to express the relationship with a simple linear shape. Therefore, in a simple linear model such as multiple regression analysis, it is considered difficult to accurately estimate the size of the damaged portion from measurement data such as vibration acceleration.

Even under such conditions, the inventors have found a method for accurately estimating the damage size by connecting multiple features of vibration acceleration and the damage size with a regression model by machine learning. This method will be described step by step below.

(Determination of learning sample)
In the method for monitoring the state of the rolling bearing 10 according to the first embodiment, first, the rolling bearing 10 (hereinafter, also referred to as “learning sample”) for generating a regression model by machine learning is determined.

Specifically, in the first embodiment, 32 rolling bearings 10 were selected as learning samples from the 33 rolling bearings 10 used in the above experiment. As will be described later, the remaining one rolling bearing 10 was used as a "test sample" for estimating the damage size using the regression model. The learning sample corresponds to one embodiment of the "first rolling bearing". The test sample corresponds to one embodiment of "second rolling bearing".

(Explanation of preprocessing of measurement data)
Next, the diagnostic unit 106 separates the vibration acceleration data of the learning sample by frequency filtering. Specifically, the vibration acceleration data is separated into data corresponding to 20 to 1000 Hz, 1000 to 5000 Hz, 5000 to 20000 Hz by a low-pass filter, a band-pass filter, and a high-pass filter.

Next, the diagnostic unit 106 divides the vibration acceleration data separated for each frequency into a plurality of segments, and calculates a feature quantity vector composed of a combination of a plurality of feature quantities for each segment.

FIG. 7 is a diagram for explaining a segment and a feature amount vector. With reference to FIG. 7, the measurement data of the time length T1 is divided into segments of the time length T2 shorter than the time length T1. For example, T2 is set to a length that is a constant multiple of the time required for one rotation of the rotation axis. For example, T2 is set at 0.6 seconds, which corresponds to the time required for the rotation axis to rotate five times. After dividing the measurement data into segments, a feature vector is generated for each of the further divided data.

The feature quantity vector treats a plurality of feature quantities as a set of vectors. When the measured data is vibration acceleration, the feature amount is, for example, the level and / or shape of the waveform such as effective value, maximum value, crest factor, kurtosis, and skewness for at least one parameter of time, frequency, and skewness. It can be a parameter representing the characteristics of. The feature amount may be processed by various frequency filters. Further, the feature amount extracted from the image data obtained by the time frequency analysis from the vibration data may be used.

As described above, by preprocessing the vibration acceleration data of the training sample, a feature vector for generating a regression model by machine learning, which will be described later, is generated. The vibration acceleration data of the test sample is also converted into a feature vector by the same preprocessing and then input to the regression model.

(Calculation of damage size)
Next, the diagnostic unit 106 obtains the damage size in each segment based on the damage size data measured in the training sample. Specifically, the damage size at the time corresponding to each segment is obtained. As a result, it is possible to obtain a combination of the feature amount vector and the damage size corresponding to each segment. That is, time-dependent data showing the relationship between the feature vector and the damage size at each stage in which the damage progresses can be obtained.

(Generation of regression model using machine learning)
FIG. 8 is a diagram for explaining the generation of a regression model by machine learning and the estimation of the damage size using the regression model.

The diagnostic unit 106 inputs the feature vector of the learning sample for each segment, outputs the damage size, and generates a regression model (first regression model) by machine learning. Hereinafter, when simply referred to as a "regression model" in the specification of the present application, the regression model by machine learning will be shown. As machine learning, for example, regression methods such as support vector machine, random forest, kernel ridge regression, deep learning regression, and combinations thereof can be used. In the first embodiment, the first regression model is performed from the initial stage of damage (that is, stage 2) in which the damage progress rate is slow, the final stage of damage (that is, stage 3) in which the damage progress rate is fast, and after the initial damage occurs. Generated for each of the whole (ie, stages 2 and 3) up to the operable limit.

As described above, the diagnostic unit 106 generates a first regression model in which the damage size can be automatically estimated by inputting the feature vector (see FIG. 8). The diagnostic unit 106 holds the first regression model in a storage unit (not shown) so that it can be read out and used at any time.

(Estimation of damage size using regression model by machine learning)
Next, a method of estimating the damage size from the vibration acceleration data of the test sample using the first regression model will be described. First, the vibration acceleration data of the test sample is converted into a feature vector by the same preprocessing as the vibration acceleration data of the training sample.

Next, the corresponding feature vector of the test sample is applied to the first regression models of the above three types, the initial stage of damage, the final stage of damage, and the whole. This gives an estimate of the damage size of the test sample as the output of the first regression model.

(Evaluation of regression model)
The results of evaluating the accuracy of the first regression model generated in the first embodiment using mathematical and statistical methods are shown below. In the first embodiment, the result using the root mean square error square root and the coefficient of determination is illustrated as the evaluation of the estimation accuracy.

FIG. 9 is a diagram for explaining the evaluation of the estimation accuracy of the first regression model. With reference to FIG. 9, in the test sample, the root mean square error square root and the coefficient of determination of the damage size estimated by the first regression model according to the first embodiment and the actually measured damage size are shown. In addition, as a comparative example, the root mean square error and the coefficient of determination when a multiple regression model is used instead of the regression model by machine learning are also shown.

The root mean square error and the coefficient of determination are calculated for each of the three types of first regression models described above. That is, the statistical value is calculated when the first regression model is generated in each of the initial stage of damage (stage 2), the final stage of damage (stage 3), and all of them (stage 2 and stage 3). There is.

The "root mean square error square root" is the square root obtained by averaging the squares of the error between the measured value of the damage size and the estimated value. That is, the root mean square error square root indicates that the smaller the value, the higher the accuracy of estimating the damage size. Referring to FIG. 9, in each of the above-mentioned three types of first regression models, the value of the root mean square error square root is smaller in the embodiment than in the comparative example. That is, it is shown that the estimation accuracy of the embodiment is higher than that of the comparative example. In addition, it is considered that the estimation accuracy is improved because the value of the root mean square error is greatly reduced, especially at the end of the damage.

The "coefficient of determination" is an index showing how well the estimated value obtained by regression analysis matches the actual measured value. Specifically, the closer the coefficient of determination is to 1, the closer the estimated value and the measured value are. Referring to FIG. 9, in each of the above-mentioned three types of first regression models, the coefficient of determination value of the embodiment is closer to 1 than that of the comparative example. That is, it is shown that the estimation accuracy of the embodiment is higher than that of the comparative example. In addition, it is considered that the estimation accuracy is improved because the value of the coefficient of determination increases significantly, especially in the final stage of injury.

As described above, in the first embodiment using the regression model by machine learning, the damage size can be estimated with higher accuracy than the comparative example using the multiple regression analysis.

FIG. 10 is an example of a flowchart for explaining the process of generating a regression model by machine learning. The flowchart shown in FIG. 10 is executed by the diagnostic unit 106.

With reference to FIG. 10, in step S01, the diagnostic unit 106 acquires the vibration acceleration data and the damage size in the training sample. Specifically, for example, the diagnostic unit 106 receives vibration acceleration data from the acceleration sensor 102 wirelessly or by wire. Further, for example, the diagnostic unit 106 receives the damage size data wirelessly or by wire from the device in which the damage size value at each measurement time measured by the user is input.

In step S02, the diagnostic unit 106 divides the vibration data into a plurality of segments. The segment is set to, for example, a length that is a constant multiple of the time required for one rotation of the rotation axis.

In step S03, the diagnostic unit 106 generates a feature amount vector for each segment. The feature amount vector treats a plurality of feature amounts of vibration data as a set of vectors.

In step S04, the diagnostic unit 106 obtains the corresponding damage size for each segment.

In step S05, the diagnostic unit 106 uses machine learning to generate a damage size regression model (first regression model) that inputs the feature vector and outputs the damage size, and ends the process. For example, the diagnostic unit 106 is configured to perform machine learning about the combination of the feature vector and the damage size in each segment of the whole (that is, stages 2 and 3) from the occurrence of the initial damage to the operable limit. Will be done.

FIG. 11 is an example of a flowchart for explaining the process of estimating the damage size using the generated regression model. The flowchart shown in FIG. 11 is executed by the diagnostic unit 106.

With reference to FIG. 11, in step S21, the diagnostic unit 106 acquires vibration data in the test sample. Specifically, for example, the diagnostic unit 106 receives vibration acceleration data from the acceleration sensor 102 wirelessly or by wire.

In step S22, the diagnostic unit 106 divides the vibration data into segments. The segment is set to, for example, a length that is a constant multiple of the time required for one rotation of the rotation axis.

In step S23, the diagnostic unit 106 generates a feature amount vector for each segment.

In step S24, the diagnostic unit 106 inputs the feature quantity vector of the test sample into the regression model (first regression model) generated by using machine learning for the relationship between the vibration data of the training sample and the damage size, and the test sample. Get an estimate of the damage size of and finish the process. This regression model is, for example, by performing machine learning about the combination of the feature vector and the damage size in each segment of the whole (that is, stages 2 and 3) from the occurrence of the above initial damage to the operable limit. It was generated. Here, the condition monitoring device 100 may be configured to notify the user of the rolling bearing 10 of the estimated damage size by using voice, visual display, or the like.

In the first embodiment, the damage size is used as an index of the damage status of the rolling bearing 10. However, if it is an index that monotonically decreases or monotonically increases with deterioration, the rated operating torque, power consumption, the rate of wear debris mixed in grease, etc. can be used as an index of the damage status instead of the damage size. ..

As described above, the condition monitoring device 100 according to the first embodiment generates a damage size regression model based on machine learning based on the relationship between the vibration data of the rolling bearing 10 for learning and the damage size, and is used for evaluation. By inputting the vibration data of the rolling bearing 10 into the regression model, the damage size is estimated. With such a configuration, the condition monitoring device 100 can estimate the size of the damaged part of the rolling bearing more accurately than the estimation accuracy by the conventional multiple regression analysis based on the vibration data that can be easily measured. it can.

(About application of rolling bearing condition monitoring device 100)
The rolling bearing condition monitoring device 100 described above can be applied to various mechanical devices, and is particularly suitable for condition monitoring of the spindle bearing of a wind power generator. That is, the spindle bearing of the wind power generator is not easy to replace, is used under relatively low speed conditions, and is often continuously used even if the bearing is damaged. For the spindle bearings of such wind power generation equipment, it is an issue to clarify the bearing replacement time due to damage.

(Structure of wind power generator)
FIG. 12 is a diagram schematically showing the configuration of a wind power generation device to which the rolling bearing condition monitoring device 100 according to the first embodiment is applied. With reference to FIG. 12, the wind power generator 210 includes a spindle 220, a blade 230, a speed increaser 240, a generator 250, a spindle bearing (hereinafter, simply referred to as “bearing”) 260, and an acceleration sensor. It includes 270 and a data processing device 280. The speed increaser 240, the generator 250, the bearing 260, the acceleration sensor 270 and the data processing device 280 are housed in the nacelle 290, and the nacelle 290 is supported by the tower 300.

The spindle 220 enters the nacelle 290, is connected to the input shaft of the speed increaser 240, and is rotatably supported by the bearing 260. Then, the main shaft 220 transmits the rotational torque generated by the blade 230 that has received the wind power to the input shaft of the speed increaser 240. The blade 230 is provided at the tip of the main shaft 220, converts wind power into rotational torque, and transmits the wind power to the main shaft 220.

The speed increaser 240 is provided between the spindle 220 and the generator 250, and increases the rotational speed of the spindle 220 to output to the generator 250. As an example, the speed increaser 240 is configured by a gear speed increase mechanism including a planetary gear, an intermediate shaft, a high speed shaft, and the like. Although not particularly shown, a plurality of bearings for rotatably supporting a plurality of shafts are also provided in the speed increaser 240. The generator 250 is connected to the output shaft of the speed increaser 240 and generates electricity by the rotational torque received from the speed increaser 240. The generator 250 is composed of, for example, an induction generator, but the type of the generator 250 is not limited to this. A bearing that rotatably supports the rotor is also provided in the generator 250.

The bearing 260 is fixed in the nacelle 290 and rotatably supports the spindle 220. The bearing 260 is a rolling bearing, and is a bearing to be monitored by the condition monitoring device 100 according to this embodiment. The bearing 260 is different from the rolling bearing 10 described below in FIG. 1 in that the inner ring is a rotating wheel and the outer ring is a stationary wheel. However, the condition monitoring device 100 according to the above embodiment is described in this way. It is also applicable to bearing 260. When the outer ring is a stationary ring, the load region is formed on the outer ring below the central axis in the vertical direction, and the initial damage occurs on the raceway surface of the inner peripheral surface of the outer ring.

The acceleration sensor 270 is a sensor for detecting the acceleration in the vertical direction (or radial direction) of the bearing 260. The acceleration sensor 270 is fixed at a place where the vibration acceleration of the bearing on the outer diameter surface of the equipment or the bearing box can be measured, for example. Further, the acceleration sensor 270 may be a sensor for detecting the vibration acceleration in the horizontal direction (or horizontal direction) of the bearing 260.

The data processing device 280 is provided in the nacelle 290 and receives the detected value of the acceleration sensor 270. Then, the data processing device 280 monitors the state of the bearing 260 according to a preset program. Specifically, the data processing device 280 detects the vibration acceleration of the bearing 260 based on the detection value of the acceleration sensor 270. The data processing device 280 generates a feature vector based on the vibration acceleration. The data processing device 280 is configured to be capable of generating a regression model by machine learning that inputs the feature amount vector and outputs the damage size based on the feature amount vector and the damage size. The data processing device 280 is configured so that the damage size can be estimated by inputting the feature amount vector based on the regression model.

The data processing device 280 realizes the function of the diagnostic unit 106 (FIG. 1) described above. Further, the data processing device 280 and the acceleration sensor 270 constitute the above-mentioned condition monitoring device 100 (FIG. 1).

As described above, the condition monitoring device 100 according to the first embodiment can estimate the damage size of the rolling bearing 10 with high accuracy based on the vibration acceleration of the rolling bearing 10 by using the regression model by machine learning. .. Therefore, it is possible to provide a condition monitoring method and a condition monitoring device capable of accurately estimating the size of the damaged portion of the rolling bearing.

[Embodiment 2]
In the first embodiment, a method of estimating the damage size of the rolling bearing 10 based on the vibration acceleration of the rolling bearing 10 has been described by using a regression model by machine learning.

In the second embodiment, a method of estimating the remaining life of the rolling bearing 10 based on the vibration acceleration of the rolling bearing 10 will be described using a regression model by machine learning. Since the condition monitoring device 100 according to the second embodiment has the same basic configuration as the condition monitoring device 100 shown in FIG. 1, the description will not be repeated.

(Determination of learning sample)
In the state monitoring method of the rolling bearing 10 according to the second embodiment, as in the state monitoring method of the rolling bearing 10 according to the first embodiment, first, the rolling bearing 10 for generating a regression model by machine learning (learning sample). ) Is determined.

In the second embodiment, specifically, 32 rolling bearings 10 were selected as learning samples from the 33 rolling bearings 10 used in the above experiment. As will be described later, the remaining one rolling bearing 10 was used as a "test sample" for estimating the remaining life using the regression model. The learning sample corresponds to one embodiment of the "first rolling bearing". The test sample corresponds to one embodiment of "second rolling bearing".

Next, the diagnostic unit 106 divides the vibration acceleration data separated for each frequency into a plurality of segments, and calculates a feature quantity vector composed of a combination of a plurality of feature quantities for each segment. Specifically, as shown in FIG. 7, the measurement data of the time length T1 is divided into segments of the time length T2 shorter than the time length T1. For example, T2 is set to a length that is a constant multiple of the time required for one rotation of the rotation axis. For example, T2 is set at 0.6 seconds, which corresponds to the time required for the rotation axis to rotate five times. After dividing the measurement data into segments, a feature vector is generated for each of the further divided data.

(Calculation of damage size)
At the same time, the diagnostic unit 106 obtains the damage size in each segment based on the damage size data measured in the training sample. Specifically, the damage size at the time corresponding to each segment is obtained. This damage size value can be used as an index of the reference of the usable limit of the rolling bearing 10 described above.

(Calculation of remaining life)
The diagnostic unit 106 obtains the remaining life in each segment. Here, the remaining life is the time from the segment to reach the standard of the usable limit. The time until the reference of the usable limit is reached is the time until the damage size exceeds a constant multiple of the interval of the rolling elements 18 in the rolling bearing 10 in this experiment. The remaining life is calculated by subtracting the operating time from the start of use of the rolling bearing 10 to the corresponding segment from the total operating time from the start of use of the rolling bearing 10 to the reaching of the limit standard in which the rolling bearing 10 can be used. Is sought after. As a result, it is possible to obtain a combination of the feature quantity vector and the remaining life corresponding to each segment. That is, time-dependent data showing the relationship between the feature vector and the remaining life at each stage in which the damage progresses can be obtained.

(Generation of regression model using machine learning)
FIG. 13 is a diagram for explaining the generation of a regression model by machine learning and the estimation of the remaining life using the regression model. The diagnostic unit 106 inputs the feature vector of the learning sample for each segment, outputs the remaining life, and generates a regression model (second regression model) by machine learning. As machine learning, for example, regression methods such as support vector machine, random forest, kernel ridge regression, deep learning regression, and combinations thereof can be used. In the second embodiment, the second regression model is used from the initial stage of damage (that is, stage 2) in which the damage progress rate is slow, the final stage of damage (that is, stage 3) in which the damage progress rate is fast, and after the initial damage occurs. Generated for each of the whole (ie, stages 2 and 3) up to the operable limit.

As described above, the diagnostic unit 106 generates a second regression model in which the remaining life can be automatically estimated by inputting the feature vector (see FIG. 13). The diagnostic unit 106 holds the second regression model in a storage unit (not shown) so that it can be read out and used at any time.

(Estimation of remaining life using a regression model by machine learning)
Next, a method of estimating the remaining life from the vibration acceleration data of the test sample using the second regression model will be described. First, the vibration acceleration data of the test sample is converted into a feature vector by the same preprocessing as the vibration acceleration data of the training sample.

Next, the corresponding feature vector of the test sample is applied to the above three types of second regression models: initial damage, final damage, and the whole. As a result, an estimated value of the remaining life of the test sample is obtained as the output of the second regression model.

(Evaluation of regression model)
The results of evaluating the accuracy of the second regression model generated in the second embodiment using mathematical and statistical methods are shown below. In the second embodiment, the result using the root mean square error square root and the coefficient of determination is illustrated as the evaluation of the estimation accuracy.

FIG. 14 is a diagram for explaining the evaluation of the estimation accuracy of the regression model. With reference to FIG. 14, in the test sample, the root mean square error square root and the coefficient of determination of the residual life estimated by the regression model (second regression model) according to the second embodiment and the measured residual life are shown. There is. Here, the actually measured remaining life or the measured value of the remaining life indicates the remaining life calculated from the total operating time that has reached the standard of the limit that can be actually used. In addition, as a comparative example, the root mean square error and the coefficient of determination when a multiple regression model is used instead of the regression model by machine learning are also shown.

The root mean square error and the coefficient of determination are calculated for each of the three types of second regression models described above. That is, the statistical value is calculated when a second regression model is generated in each of the initial stage of damage (stage 2), the final stage of damage (stage 3), and all of them (stage 2 and stage 3). There is.

The "root mean square error square root" is the square root obtained by averaging the square of the error between the measured value and the estimated value of the remaining life. That is, the root mean square error square root indicates that the smaller the value, the higher the accuracy of estimating the remaining life. Referring to FIG. 14, in each of the above-mentioned three types of second regression models, the value of the root mean square error square root is smaller in the embodiment than in the comparative example. That is, it is shown that the estimation accuracy of the embodiment is higher than that of the comparative example. In addition, it is considered that the estimation accuracy is improved because the value of the root mean square error is greatly reduced, especially at the end of the damage.

The "coefficient of determination" is an index showing how well the estimated value obtained by regression analysis matches the actual measured value. Specifically, the closer the coefficient of determination is to 1, the closer the estimated value and the measured value are. Referring to FIG. 14, in each of the above-mentioned three types of second regression models, the coefficient of determination value of the embodiment is closer to 1 than that of the comparative example. That is, it is shown that the estimation accuracy of the embodiment is higher than that of the comparative example. In addition, it is considered that the estimation accuracy is improved because the value of the coefficient of determination increases significantly, especially at the end of injury.

As described above, in the second embodiment using the regression model by machine learning, the remaining life can be estimated with higher accuracy than the comparative example using the multiple regression analysis.

FIG. 15 is an example of a flowchart for explaining the process of generating a regression model by machine learning. The flowchart shown in FIG. 15 is executed by the diagnostic unit 106.

With reference to FIG. 15, in step S31, the diagnostic unit 106 acquires the vibration acceleration data in the learning sample and the total operating time when the bearing 10 reaches the usable limit reference. Specifically, for example, the diagnostic unit 106 receives vibration acceleration data from the acceleration sensor 102 wirelessly or by wire. Further, for example, the diagnostic unit 106 receives the data of the total operating time wirelessly or by wire from the device in which the total operating time is input by the user.

In step S32, the diagnostic unit 106 divides the vibration data into a plurality of segments. The segment is set to, for example, a length that is a constant multiple of the time required for one rotation of the rotation axis.

In step S33, the diagnostic unit 106 generates a feature amount vector for each segment. The feature amount vector treats a plurality of feature amounts of vibration data as a set of vectors.

In step S34, the diagnostic unit 106 obtains the remaining life corresponding to each segment. Specifically, the diagnostic unit 106 calculates the remaining life by subtracting the operation time corresponding to the segment from the total operation time.

In step S35, the diagnostic unit 106 uses machine learning to generate a regression model of the remaining life (second regression model) that inputs the feature vector and outputs the remaining life, and ends the process. For example, the diagnostic unit 106 is configured to perform machine learning about the combination of the feature vector and the remaining life in each segment of the whole (that is, stages 2 and 3) from the occurrence of the initial damage to the operable limit. Will be done.

FIG. 16 is an example of a flowchart for explaining a process of estimating the remaining life using the generated regression model. The flowchart shown in FIG. 16 is executed by the diagnostic unit 106.

With reference to FIG. 16, in step S41, the diagnostic unit 106 acquires vibration data in the test sample. Specifically, for example, the diagnostic unit 106 receives vibration acceleration data from the acceleration sensor 102 wirelessly or by wire.

In step S42, the diagnostic unit 106 divides the vibration data into segments. The segment is set to, for example, a length that is a constant multiple of the time required for one rotation of the rotation axis.

In step S43, the diagnostic unit 106 generates a feature amount vector for each segment.

In step S44, the diagnostic unit 106 inputs the feature quantity vector of the test sample into the regression model (second regression model) generated by using machine learning for the relationship between the vibration data of the training sample and the remaining life, and the test sample. The estimated value of the remaining life of is obtained, and the process is terminated. The second regression model is, for example, machine learning about the combination of the feature vector and the remaining life in each segment of the whole (that is, stages 2 and 3) from the occurrence of the above initial damage to the operable limit. It is generated by doing. Here, the condition monitoring device 100 may be configured to notify the user of the rolling bearing 10 of the estimated remaining life by using voice, visual display, or the like.

As described above, the condition monitoring device 100 according to the second embodiment generates a regression model of the remaining life based on machine learning based on the relationship between the vibration data of the rolling bearing 10 for learning and the remaining life, and is used for evaluation. By inputting the vibration data of the rolling bearing 10 into the regression model, the remaining life is estimated. With such a configuration, the condition monitoring device 100 can estimate the remaining life of the rolling bearing more accurately than the estimation accuracy by the conventional multiple regression analysis based on the vibration data that can be easily measured.

(Structure of wind power generator)
FIG. 17 is a diagram schematically showing the configuration of a wind power generation device to which the rolling bearing condition monitoring device 100 according to the second embodiment is applied. The wind power generation device 210 according to the second embodiment replaces the data processing device 280 in the wind power generation device 210 shown in FIG. 12 with the data processing device 280A. The description of the common parts with the first embodiment will not be repeated.

The data processing device 280A is provided in the nacelle 290 and receives the detected value of the acceleration sensor 270. Then, the data processing device 280A monitors the state of the bearing 260 according to a preset program. Specifically, the data processing device 280A detects the vibration acceleration of the bearing 260 based on the detection value of the acceleration sensor 270. The data processing device 280A generates a feature vector based on the vibration acceleration. The data processing device 280 is configured to be capable of generating a regression model by machine learning based on the feature amount vector and the remaining life, with the feature amount vector as an input and the remaining life as an output. The data processing device 280 is configured so that the remaining life can be estimated by inputting the feature amount vector based on the regression model.

The data processing device 280A realizes the function of the diagnostic unit 106 (FIG. 1). Further, the data processing device 280A and the acceleration sensor 270 constitute the condition monitoring device according to the second embodiment.

As described above, the condition monitoring device 100 according to the second embodiment can estimate the remaining life of the rolling bearing 10 with high accuracy based on the vibration acceleration of the rolling bearing 10 by using the regression model by machine learning. .. Therefore, it is possible to provide a condition monitoring method and a condition monitoring device capable of accurately estimating the remaining life of the rolling bearing.

In addition, the embodiments disclosed this time should be considered to be exemplary in all respects and not restrictive. The scope of the present invention is shown by the 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 claims.

10,260 bearing, 12 inner ring, 14 shaft body, 16 outer ring, 18 rolling element, 20 raceway surface, 100 condition monitoring device, 102,270 acceleration sensor, 106 diagnostic unit, 210 wind power generator, 220 spindle, 230 blade, 240 Accelerometer, 250 generator, 280, 280A data processor, 290 nacelle, 300 tower, D1, D2, D3 damaged.

Claims

The step of generating a regression model by machine learning based on the vibration data of the first rolling bearing,
A step of monitoring the state of the second rolling bearing using the regression model is provided.
Each of the first rolling bearing and the second rolling bearing
Rotating wheel and
With a stationary wheel
A plurality of rolling elements arranged between the rotating wheel and the stationary wheel and moving on the raceway surface of the stationary wheel as the rotating wheel rotates are provided.
Let the length in the circumferential direction of the damaged portion formed on the raceway surface of the stationary wheel of the rolling bearing be the damage size.
The step of generating the regression model is
The step of acquiring the vibration data of the first rolling bearing and the damage size of the first rolling bearing, and
A step of generating a feature amount vector of the first rolling bearing composed of a feature amount of vibration data of the first rolling bearing, and a step of generating the feature amount vector of the first rolling bearing.
A step of generating a first regression model for estimating the damage size from the feature vector by machine learning the feature vector of the first rolling bearing and the damage size of the first rolling bearing. Including
The step of monitoring the condition of the second rolling bearing is
The step of acquiring the vibration data of the second rolling bearing, and
A step of generating a feature amount vector of the second rolling bearing composed of the feature amount of the vibration data of the second rolling bearing, and a step of generating the feature amount vector of the second rolling bearing.
A method for monitoring the state of a rolling bearing, which comprises a step of estimating the damage size of the second rolling bearing from the feature quantity vector of the second rolling bearing using the first regression model.
The step of generating the regression model further includes a step of dividing the vibration data of the first rolling bearing into segments having a length constant times the time required for one rotation of the rotating wheel.
The feature vector of the first rolling bearing is generated for each segment.
The first regression model is generated by performing machine learning between the feature amount vector of the first rolling bearing generated for each segment and the damage size of the first rolling bearing, according to claim 1. The method for monitoring the condition of rolling bearings described.
The method for monitoring the state of a rolling bearing according to claim 2, wherein each of the vibration data of the first and second rolling bearings includes data of vibration acceleration.
The vibration data of the first rolling bearing and the damage size of the first rolling bearing are the time until the damage size of the first rolling bearing becomes a constant multiple of the rolling element spacing, or the first one. The state of a rolling bearing according to any one of claims 1 to 3, wherein a value of a predetermined feature amount that changes reflecting the damage size of the rolling bearing is acquired over a period of time until it deviates from a predetermined allowable range. Monitoring method.
Assuming that the time from the acquisition of vibration data until the rolling bearing reaches the usage limit is the remaining life.
The step of generating the regression model further
The step of calculating the remaining life of the first rolling bearing and
A step of generating a second regression model for estimating the remaining life from the feature vector by machine learning the feature vector of the first rolling bearing and the remaining life of the first rolling bearing. Including
The step of monitoring the condition of the second rolling bearing is
The condition monitoring of the rolling bearing according to claim 1, further comprising a step of estimating the remaining life of the second rolling bearing from the feature quantity vector of the second rolling bearing using the second regression model. Method.
The step of generating the regression model further includes a step of dividing the vibration data of the first rolling bearing into segments having a length constant times the time required for one rotation of the rotating wheel.
The feature vector of the first rolling bearing is generated for each segment.
The second regression model is generated by performing machine learning of the feature amount vector of the first rolling bearing generated for each segment and the remaining life of the first rolling bearing, according to claim 5. The method for monitoring the condition of rolling bearings described.
The method for monitoring the state of a rolling bearing according to claim 6, wherein each of the vibration data of the first and second rolling bearings includes data of vibration acceleration.
The step of generating the regression model further includes the step of acquiring the damage size of the first rolling bearing.
The time until the rolling bearing reaches the usage limit reflects the time until the damage size of the first rolling bearing becomes a constant multiple of the rolling element spacing, or the damage size of the first rolling bearing. The method for monitoring the state of a rolling bearing according to any one of claims 5 to 7, which is the time required for the value of the predetermined feature amount to change to deviate from the predetermined allowable range.
A condition monitoring device that monitors the condition of rolling bearings by using the condition monitoring method according to any one of claims 1 to 8.