Abnormality detection device for rolling bearing, and method for generating logistic regression model

Abstract

This abnormality detection device for a rolling bearing is provided with a detection device body. The detection device body computes, from vibration data acquired from an acceleration sensor provided to the rolling bearing, an overall value and the peak height of a frequency, in a frequency waveform, which arises from damage to the rolling bearing. The detection device body also computes, on the basis of a logistic regression model and a plurality of reference values including the overall value and the peak height, an abnormality probability that represents the probability of damage occurring inside the rolling bearing.

Description

Rolling bearing abnormality detection device and logistic regression model generation method

[0001] The present disclosure relates to an abnormality detection device for a rolling bearing and a method for generating a logistic regression model.

[0002] In a conventional condition monitoring method, measurement data acquired from a sensor installed on an object is filtered to calculate multiple feature quantities. Then, machine learning is performed using the multiple feature quantities as learning data to determine an algorithm for determining whether the object is abnormal (see, for example, Patent Document 1).

[0003] JP 2019-45484 A

[0004] In the conventional condition monitoring methods described above, only measurement data acquired when the object is normal is used as learning data for machine learning. Therefore, by the time an abnormality is determined, damage to the object may have progressed. For example, in rolling bearings used at relatively low rotation speeds, such as bearings in elevator hoists, vibrations are small in the early stages of damage, and the bearings are not determined to be abnormal. By the time an abnormality is detected, the damage may have progressed significantly.

[0005] The present disclosure has been made to solve the above-mentioned problems, and aims to provide a rolling bearing abnormality detection device that can detect abnormalities in rolling bearings at an earlier stage, and a method for generating a logistic regression model.

[0006] The rolling bearing abnormality detection device according to the present disclosure comprises a detection device main body that calculates an overall value and the peak height of a frequency in a frequency waveform that is caused by damage to the rolling bearing from vibration data acquired from an acceleration sensor provided in the rolling bearing, and calculates an abnormality probability, which is the probability that damage has occurred inside the rolling bearing, based on a plurality of reference values ​​including the overall value and the peak height and a logistic regression model, where the logistic regression model is a logistic function generated by performing logistic regression analysis using a plurality of indices including the overall value and the peak height from training data of a plurality of samples that includes normal data, which is vibration data obtained by rotating a plurality of rolling bearings in a normal state, and abnormal data, which is vibration data obtained by rotating a plurality of rolling bearings in a damaged state. The method for generating a logistic regression model according to the present disclosure includes the steps of acquiring training data of a plurality of samples including normal data, which is vibration data obtained by rotating a plurality of rolling bearings in a normal state, and abnormal data, which is vibration data obtained by rotating a plurality of rolling bearings in a damaged state, classifying the training data into two values: normal data and abnormal data, and selecting a plurality of indices from the binary-classified training data, performing logistic regression analysis, and deriving, as a logistic regression model, a logistic function for calculating the abnormality probability, which is the probability that damage has occurred inside the rolling bearing.

[0007] According to the present disclosure, abnormalities in rolling bearings can be detected at an earlier stage.

[0008] 1. A configuration diagram showing a rolling bearing according to embodiment 1. A block diagram showing the rolling bearing abnormality detection device of FIG. 1. A graph showing an example of vibration data acquired by the data acquisition unit of FIG. 2. A graph showing an example of a frequency waveform output by the envelope analysis unit of FIG. 2. An explanatory diagram showing a method of calculating peak heights in the peak height calculation unit of FIG. 2. A flowchart showing rolling bearing state determination processing by the detection device main body of FIG. 2. A flowchart showing model generation processing by the detection device main body of FIG. 2. A table showing an example of learning data used in the model generation processing of embodiment 1. A flowchart showing model generation processing by the detection device main body of embodiment 2. A block diagram showing a rolling bearing abnormality detection device according to embodiment 3. A flowchart showing model generation processing by the detection device main body of FIG. 10. A table showing an example of learning data used in the model generation processing of embodiment 3. A flowchart showing model generation processing by the detection device main body of embodiment 4. A flowchart showing model generation processing by the detection device main body of embodiment 5. A flowchart showing model generation processing by the detection device main body of embodiment 6. A block diagram showing main parts of a rolling bearing abnormality detection device according to embodiment 7. A configuration diagram showing a first example of a processing circuit that realizes each function of the detection device main body of embodiments 1 to 7. FIG. 10 is a configuration diagram showing a second example of a processing circuit that realizes each function of the detection device main body of the first to seventh embodiments.

[0009] Hereinafter, embodiments will be described with reference to the drawings. Embodiment 1. Fig. 1 is a structural diagram showing a rolling bearing according to embodiment 1. In the figure, rolling bearing 11 has an outer ring 12, an inner ring 13, and a plurality of rolling elements 14. Outer ring 12 is incorporated into housing 15. Inner ring 13 is disposed inside outer ring 12.

[0010] The plurality of rolling elements 14 are interposed between the outer ring 12 and the inner ring 13. The plurality of rolling elements 14 are also arranged at equal intervals from one another in the circumferential direction of the inner ring 13. Each of the rolling elements 14 is a ball or roller.

[0011] The inner ring 13 is attached to the rotary shaft 16. The inner ring 13 rotates together with the rotary shaft 16 around the axis of the rotary shaft 16.

[0012] Although the rolling bearing 11 of the first embodiment is of a type in which the inner ring 13 rotates, it may be of a type in which the outer ring 12 rotates.

[0013] When the inner ring 13 rotates, vibrations occur in the rolling bearing 11. An acceleration sensor 17 is provided in the housing 15. The acceleration sensor 17 generates a signal corresponding to the vibrations of the rolling bearing 11.

[0014] Fig. 2 is a block diagram showing an abnormality detection device for the rolling bearing 11 of Fig. 1. The abnormality detection device 20 has a detection device main body 21 and a monitor 22.

[0015] The detection device main body 21 has the following functional blocks: a vibration data acquisition unit 23, a filter processing unit 24, an overall value calculation unit 25, an envelope analysis unit 26, a peak height calculation unit 27, a multiplication unit 28, a model generation unit 29, a model memory unit 30, a probability calculation unit 31, a judgment unit 32, and a display unit 33.

[0016] The vibration data acquisition unit 23 acquires vibration data output from the acceleration sensor 17 when the rotation shaft 16 is rotating at a constant speed.

[0017] Fig. 3 is a graph showing an example of vibration data acquired by the vibration data acquisition unit 23. Fig. 2 shows the change in vibration acceleration of the rolling bearing 11 over time.

[0018] The filter processing unit 24 generates noise-removed data, which is data from which noise has been removed, by performing filtering on the vibration data acquired by the vibration data acquisition unit 23. Examples of filtering include band-pass filtering, low-pass filtering, and high-pass filtering.

[0019] The overall value calculation unit 25 calculates an overall value from the noise-removed data generated by the filter processing unit 24. The overall value is the sum of the magnitudes of the frequency components in all frequency bands after FFT (Fast Fourier Transform) analysis.

[0020] If the circularity of the rolling bearing 11 deteriorates due to wear or the like, or if damage occurs inside the rolling bearing 11, the overall value of the vibration of the rolling bearing 11 increases. Furthermore, after some time has passed since damage occurred inside the rolling bearing 11, the corners of the damaged area become rounded and the peak becomes smaller, but the overall value often remains large.

[0021] The envelope analysis unit 26 performs envelope processing on the noise-removed data generated by the filter processing unit 24, and outputs a frequency waveform that has been Fourier transformed.

[0022] Figure 4 is a graph showing an example of a frequency waveform output by the envelope analysis unit 26. As shown in Figure 4, when damage occurs inside the rolling bearing 11, peaks occur at the bearing damage frequency and at frequencies that are integer multiples of the bearing damage frequency. The bearing damage frequency is a frequency caused by damage inside the rolling bearing 11, and is determined by the dimensions of the rolling bearing 11, the number of rolling elements 14, etc.

[0023] The peak height calculation unit 27 calculates the peak height of the bearing damage frequency in the output frequency waveform.

[0024] 5 is an explanatory diagram showing a method for calculating the peak height in the peak height calculation unit 27. When damage occurs inside the rolling bearing 11, a peak occurs at the bearing damage frequency. Furthermore, depending on the sampling frequency, a peak may occur at a frequency close to the bearing damage frequency.

[0025] In calculating the peak height, multiple sampling points are set on both the lower and higher sides of the frequency at which the peak occurs, i.e., the peak frequency. The lowest point (X1, Y1) is selected from the multiple sampling points on the lower side of the peak frequency. The lowest point (X2, Y2) is also selected from the multiple sampling points on the higher side of the peak frequency.

[0026] The peak height calculation unit 27 uses the lowest point (X1, Y1) and the lowest point (X2, Y2) to calculate the height H to the top (Xp, Yp) of the peak frequency using one of the following three equations, where A and B are constants.

[0027] H=[(Yp-Y1)+(Yp-Y2)] / 2

[0028] H=A×[(Yp-Y1)+(Yp-Y2)] / 2+B

[0029] H= (Yp-Y1)-[(Y2-Y1) / (X2-X1)]×(X2-X1)

[0030] The multiplication unit 28 calculates a value obtained by multiplying the output value of the overall value calculation unit 25 by the output value of the peak height calculation unit 27, that is, the value of the overall value x the peak height.

[0031] The model generation unit 29 performs machine learning on normal data and abnormal data to generate a logistic regression model. The machine learning is, for example, supervised learning. The method for generating the logistic regression model will be described in detail later.

[0032] The model storage unit 30 stores the logistic regression model generated by the model generation unit 29 .

[0033] The probability calculation unit 31 calculates the abnormality probability of the rolling bearing 11 based on a plurality of reference values ​​and a logistic regression model. The abnormality probability of the rolling bearing 11 is the probability that damage has occurred inside the rolling bearing 11 that is the inspection target. In the probability calculation unit 31 of the first embodiment, the overall value and the peak height are used as the plurality of reference values.

[0034] The determination unit 32 compares the abnormality probability calculated by the probability calculation unit 31 with a threshold value for the abnormality probability to determine whether or not there is an abnormality in the rolling bearing 11. The threshold value for the abnormality probability is set in advance in the determination unit 32.

[0035] If the abnormality probability is equal to or greater than the threshold, the determination unit 32 determines that an abnormality exists, that is, that damage has occurred in the rolling bearing 11. If the abnormality probability is less than the threshold, the determination unit 32 determines that the rolling bearing 11 is in a normal state.

[0036] The display unit 33 displays the determination result by the determination unit 32 on the monitor 22. The display unit 33 may also display the abnormality probability calculated by the probability calculation unit 31 on the monitor 22.

[0037] Fig. 6 is a flowchart showing the state determination process of the rolling bearing 11 by the detection device main body 21 in Fig. 2. The detection device main body 21 executes the state determination process when the rotating shaft 16 is rotating at a constant speed.

[0038] When the state determination process is started, the detection device main body 21 acquires vibration data from the acceleration sensor 17 in step S101.

[0039] Next, in step S102, the detection device main body 21 performs a filter process on the acquired vibration data, and then in step S103, the detection device main body 21 calculates an overall value.

[0040] Next, in step S104, the detection device main body 21 performs envelope analysis to obtain a Fourier-transformed frequency waveform. Then, in step S105, the detection device main body 21 calculates the peak height. In step S106, the detection device main body 21 calculates the value of the overall value multiplied by the peak height.

[0041] Thereafter, in step S107, the detection device main body 21 calculates the abnormality probability. Then, in step S108, the detection device main body 21 compares the abnormality probability with a threshold value to determine whether or not there is an abnormality in the rolling bearing 11. Next, in step S109, the detection device main body 21 displays the determination result on the monitor 22, and ends the state determination process.

[0042] 7 is a flowchart showing the model generation process performed by the detection device main body 21 in FIG. 2. The model generation process is a process for generating a logistic regression model. When the detection device main body 21 starts the model generation process, in step S201, the detection device main body 21 acquires training data for supervised learning.

[0043] In supervised learning, learning data of a plurality of samples is used. The plurality of samples includes a plurality of rolling bearings in a normal state and a plurality of rolling bearings in an abnormal state, i.e., a damaged state. The learning data of the plurality of samples includes normal data and abnormal data. The normal data is vibration data obtained by rotating a plurality of rolling bearings in a normal state. The abnormal data is vibration data obtained by rotating a plurality of rolling bearings in a damaged state.

[0044] After acquiring the learning data, the detection device main body 21 classifies the learning data of a plurality of samples into binary values ​​in step S202, with normal data being "0" and abnormal data being "1".

[0045] Fig. 8 is a table showing an example of training data used in the model generation process of Embodiment 1. Fig. 8 shows training data after binary classification.

[0046] In the supervised learning of the first embodiment, for each sample, the overall value, the peak height, and the value of the overall value x the peak height are used as input data, and the judgment result of whether or not the sample is normal is used as output data.

[0047] Thereafter, in step S203, the detection device main body 21 performs logistic regression analysis. Logistic regression analysis is a statistical method that uses multiple indices to predict the probability of a binary outcome. In this example, the detection device main body 21 selects two indices, the overall value and the peak height, from the binary-classified learning data.

[0048] Then, in step S204, the detection device main body 21 derives the following logistic function P as a logistic regression model, and ends the process.

[0049] P = 1 / (1 + e －ｚ )

[0050] z = A1 x overall value + A2 x peak height + A3

[0051] Note that A1, A2, and A3 are constants, which are derived by logistic regression analysis.

[0052] The output of the logistic function P is a value between 0 and 1, so the abnormality probability of the rolling bearing 11 can be determined from a plurality of reference values ​​calculated from the vibration data of the rolling bearing 11.

[0053] The method for generating a logistic regression model in the first embodiment includes a first step, a second step, and a third step.

[0054] The first step is to acquire training data of multiple samples including normal data and abnormal data. The second step is to classify the training data into binary categories: normal data and abnormal data. The third step is to select multiple indicators from the binary-classified training data, perform logistic regression analysis, and derive a logistic function.

[0055] In such an abnormality detection device 20, the detection device main body 21 acquires vibration data from the acceleration sensor 17 and calculates an overall value and a peak height. The detection device main body 21 also calculates an abnormality probability of the rolling bearing 11 based on a plurality of reference values ​​including the overall value and the peak height and a logistic regression model.

[0056] This makes it possible to more quickly detect an abnormality in the rolling bearing 11. Furthermore, even if the rolling bearing 11 is a rolling bearing used at a relatively low rotation speed, such as a bearing for an elevator hoisting machine, it is possible to more quickly detect an abnormality in the rolling bearing 11. This makes it possible to shorten the period during which the elevator is out of service.

[0057] A threshold value for the probability of abnormality is set in the detection device main body 21. The detection device main body 21 then compares the calculated probability of abnormality with the threshold value to determine whether or not there is an abnormality in the rolling bearing 11. This makes it possible to detect an abnormality in the rolling bearing 11 more quickly.

[0058] Furthermore, the vibration data acquisition unit 23 acquires vibration data from the acceleration sensor 17. The filter processing unit 24 generates noise-removed data by performing filter processing on the vibration data acquired by the vibration data acquisition unit 23. The overall value calculation unit 25 calculates an overall value from the noise-removed data. The envelope analysis unit 26 performs envelope processing on the noise-removed data and outputs a Fourier-transformed frequency waveform. The peak height calculation unit 27 calculates peak heights from the frequency waveform output by the envelope analysis unit 26.

[0059] Therefore, the overall value and peak height can be easily calculated by the detection device main body 21.

[0060] In addition, in the method for generating a logistic regression model, training data of a plurality of samples including normal data and abnormal data is acquired, and the training data is binary-classified into normal data and abnormal data. Then, a plurality of indices are selected from the binary-classified training data, and a logistic regression analysis is performed, and a logistic function for calculating the abnormality probability is derived as the logistic regression model.

[0061] Therefore, by using the generated logistic regression model, an abnormality in the rolling bearing 11 can be detected more quickly.

[0062] The multiple indices include an overall value and a peak height, which is calculated from a frequency waveform obtained by performing frequency analysis on the filtered data through envelope processing.

[0063] Therefore, a more effective logistic regression model can be easily generated, and by using this logistic regression model, abnormalities in the rolling bearing 11 can be detected at an earlier stage.

[0064] In the first embodiment, the value of the overall value multiplied by the peak height is not selected as an index, and therefore may be excluded from the input data for supervised learning. Also, in the first embodiment, the value of the overall value multiplied by the peak height is not used as a reference value, and therefore the multiplication unit 28 may be omitted.

[0065] 9 is a flowchart showing a model generation process by the detection device main body 21 according to embodiment 2. In FIG. 9, the processes in steps S201 and S202 are the same as those in embodiment 1.

[0066] The detection device main body 21 of the second embodiment performs a binary classification of data of a plurality of samples, and then performs a logistic regression analysis in step S205. The detection device main body 21 of the second embodiment selects three indices from the binary-classified learning data: an overall value, a peak height, and a value obtained by multiplying the overall value and the peak height.

[0067] Then, in step S206, the detection device main body 21 derives the following logistic function P as a logistic regression model, and ends the process.

[0068] P = 1 / (1 + e －ｚ )

[0069] z = A1 × overall value + A2 × peak height + A3 × overall value × peak height + A4

[0070] It should be noted that A1, A2, A3, and A4 are each constants derived by logistic regression analysis.

[0071] In this way, the multiple indexes in embodiment 2 include the overall value, the peak height, and the value of the overall value x the peak height. Therefore, the multiple reference values ​​in embodiment 2 also include the overall value, the peak height, and the value of the overall value x the peak height.

[0072] Other configurations and operations in the second embodiment are the same as those in the first embodiment.

[0073] In the anomaly detection device 20 and the method for generating a logistic regression model, the multiple indices include the value of the overall value x the peak height. Therefore, even if, for example, damage to the rolling bearing 11 causes an increase in only one of the overall value or the peak height, an abnormality in the rolling bearing 11 can be detected with higher accuracy.

[0074] Embodiment 3 Next, Fig. 10 is a block diagram showing a rolling bearing abnormality detection device according to embodiment 3. In addition to the same configuration as in Fig. 2, the detection device main body 21 has, as functional blocks, an operating data acquisition unit 34 and a total rotation speed calculation unit 35.

[0075] The operation data acquisition unit 34 acquires operation data related to the rolling bearing 11 from the control device 18. The operation data includes the operation time of the rolling bearing 11, the rotation speed of the rolling bearing 11, the operation pattern of the rolling bearing 11, etc. For example, if the rolling bearing 11 is a bearing for an elevator hoisting machine, the control device 18 is an elevator control device that controls the elevator hoisting machine.

[0076] As specified in the international standard ISO218 as a basic rating life L10 with a reliability of 90%, the operating data of the rolling bearing 11 is also an important factor in determining the condition of the rolling bearing 11.

[0077] The total rotation speed calculation unit 35 calculates the total rotation speed of the rolling bearing 11 , i.e., the total rotation speed of the rotating shaft 16 , based on the operating data acquired by the operating data acquisition unit 34 .

[0078] The probability calculation unit 31 of the third embodiment calculates the abnormality probability of the rolling bearing 11 based on a plurality of reference values ​​and a logistic regression model. The plurality of reference values ​​used are the output values ​​from the overall value calculation unit 25, the peak height calculation unit 27, and the operating data acquisition unit 34.

[0079] Fig. 11 is a flowchart showing the model generation process by the detection device main body 21 of Fig. 10. When the detection device main body 21 starts the model generation process, it acquires learning data in step S301.

[0080] After acquiring the learning data, the detection device main body 21 classifies the data of the multiple samples into binary values ​​in step S302, with normal data being "0" and abnormal data being "1".

[0081] Fig. 12 is a table showing an example of training data used in the model generation process of embodiment 3. Fig. 12 shows training data after binary classification.

[0082] In the supervised learning of the third embodiment, for each sample, the overall value, the peak height, the value of the overall value x the peak height, the total number of rotations, and the operating time are used as input data, and the judgment result of whether or not the sample is normal is used as output data.

[0083] Thereafter, the detection device main body 21 performs a logistic regression analysis in step S303. The detection device main body 21 of the third embodiment selects three indices, namely, the overall value, the peak height, and the operating time, from the binary-classified learning data.

[0084] Then, in step S304, the detection device main body 21 derives the following logistic function P as a logistic regression model, and ends the process.

[0085] P = 1 / (1 + e －ｚ )

[0086] z = A1 x overall value + A2 x peak height + A3 x operating time + A4

[0087] It should be noted that A1, A2, A3, and A4 are each constants derived by logistic regression analysis.

[0088] In this way, the overall value, the peak height, and the operating time are included in the multiple indexes in embodiment 3. Therefore, the multiple reference values ​​in embodiment 3 also include the overall value, the peak height, and the operating time of rolling bearing 11.

[0089] Other configurations and operations in the third embodiment are the same as those in the first embodiment.

[0090] In the abnormality detection device 20, the operating time is included in each of the multiple indexes and multiple reference values, so that abnormalities in the rolling bearing 11 can be detected with higher accuracy.

[0091] Furthermore, in the method for generating a logistic regression model, the operating time is included in the multiple indices, so that by using the generated logistic regression model, it is possible to detect an abnormality in the rolling bearing 11 with higher accuracy.

[0092] In the third embodiment, the value of the product of the overall value and the peak height and the total number of rotations are not selected as indexes, and therefore may be excluded from the input data for supervised learning. In the third embodiment, the value of the product of the overall value and the peak height and the total number of rotations are not used as reference values, and therefore the multiplication unit 28 and the total number of rotations calculation unit 35 may be omitted.

[0093] 13 is a flowchart showing a model generation process by the detection device main body 21 according to embodiment 4. In FIG. 13, the processes in steps S301 and S302 are the same as those in embodiment 3.

[0094] The detection device main body 21 of the fourth embodiment performs a binary classification of the data of a plurality of samples, and then performs a logistic regression analysis in step S305. The detection device main body 21 of the fourth embodiment selects three indices from the binary-classified learning data: the overall value, the peak height, and the total number of rotations.

[0095] Then, in step S306, the detection device main body 21 derives the following logistic function P as a logistic regression model, and ends the process.

[0096] P = 1 / (1 + e －ｚ )

[0097] z = A1 x overall value + A2 x peak height + A3 x total number of rotations + A4

[0098] It should be noted that A1, A2, A3, and A4 are each constants derived by logistic regression analysis.

[0099] In this way, the multiple indicators in embodiment 4 include the overall value, the peak height, and the total rotation speed. Therefore, the multiple reference values ​​in embodiment 4 also include the overall value, the peak height, and the total rotation speed of the rolling bearing 11.

[0100] Other configurations and operations in the fourth embodiment are the same as those in the third embodiment.

[0101] In the abnormality detection device 20, the total rotation speed is included in each of the multiple indexes and multiple reference values, so that abnormalities in the rolling bearing 11 can be detected with higher accuracy.

[0102] Furthermore, in the method for generating a logistic regression model, the total rotation speed is included in the multiple indicators, so that by using the generated logistic regression model, it is possible to detect an abnormality in the rolling bearing 11 with higher accuracy.

[0103] Furthermore, even when the rolling bearing 11 is operated not in a continuous manner but in a programmed manner including acceleration / deceleration and stopping, abnormalities in the rolling bearing 11 can be detected more accurately.

[0104] In the fourth embodiment, the value of the overall value multiplied by the peak height and the driving time are not selected as indices, and therefore may be excluded from the input data for supervised learning. In the fourth embodiment, the value of the overall value multiplied by the peak height is not used as a reference value, and therefore the multiplication unit 28 may be omitted.

[0105] 14 is a flowchart showing a model generation process by the detection device main body 21 according to embodiment 5. In FIG. 14, the processes in steps S301 and S302 are the same as those in embodiment 3.

[0106] The detection device main body 21 of the fifth embodiment classifies the data of a plurality of samples into binary values, and then performs logistic regression analysis in step S307. The detection device main body 21 of the fifth embodiment selects four indices from the binary-classified learning data: the overall value, the peak height, the product of the overall value and the peak height, and the driving time.

[0107] Then, in step S308, the detection device main body 21 derives the following logistic function P as a logistic regression model, and ends the process.

[0108] P = 1 / (1 + e －ｚ )

[0109] z = A1 x overall value + A2 x peak height + A3 x overall value x peak height + A4 x operating time + A5

[0110] It should be noted that A1, A2, A3, A4, and A5 are constants, which are derived by logistic regression analysis.

[0111] In this way, the multiple indexes in embodiment 5 include the overall value, the peak height, the value of the overall value × the peak height, and the operating time. Therefore, the multiple reference values ​​in embodiment 4 also include the overall value, the peak height, the value of the overall value × the peak height, and the operating time of rolling bearing 11.

[0112] Other configurations and operations in the fifth embodiment are the same as those in the third embodiment.

[0113] In the abnormality detection device 20, each of the multiple indexes and multiple reference values ​​includes the overall value x peak height value and the operating time, which allows for even more accurate detection of abnormalities in the rolling bearing 11.

[0114] Furthermore, in the method for generating a logistic regression model, the multiple indices include the overall value x peak height value and the operating time, so that by using the generated logistic regression model, it is possible to detect abnormalities in the rolling bearing 11 with even greater accuracy.

[0115] In the fifth embodiment, the total number of rotations is not selected as an index, and therefore may be excluded from the input data for supervised learning. In the fifth embodiment, the total number of rotations is not used as a reference value, and therefore the total number of rotations calculation unit 35 may be omitted.

[0116] 15 is a flowchart showing the model generation process by the detection device main body 21 according to embodiment 6. In FIG. 15, the processes in steps S301 and S302 are the same as those in embodiment 3.

[0117] The detection device main body 21 of the sixth embodiment performs a binary classification of the data of the plurality of samples, and then performs a logistic regression analysis in step S309. The detection device main body 21 of the sixth embodiment selects four indices from the binary-classified learning data: the overall value, the peak height, the product of the overall value and the peak height, and the total number of rotations.

[0118] Then, in step S310, the detection device main body 21 derives the following logistic function P as a logistic regression model, and ends the process.

[0119] P = 1 / (1 + e －ｚ )

[0120] z = A1 x overall value + A2 x peak height + A3 x overall value x peak height + A4 x total number of rotations + A5

[0121] It should be noted that A1, A2, A3, A4, and A5 are constants, which are derived by logistic regression analysis.

[0122] In this way, the multiple indexes in embodiment 6 include the overall value, the peak height, the value of the overall value x the peak height, and the total rotation speed. Therefore, the multiple reference values ​​in embodiment 4 also include the overall value, the peak height, the value of the overall value x the peak height, and the total rotation speed of the rolling bearing 11.

[0123] Other configurations and operations in the sixth embodiment are the same as those in the third embodiment.

[0124] In the abnormality detection device 20, each of the multiple indexes and multiple reference values ​​includes the value of the overall value x the peak height and the total rotation speed, which allows for even more accurate detection of abnormalities in the rolling bearing 11.

[0125] Furthermore, in the method for generating a logistic regression model, the multiple indices include the overall value x peak height value and the total rotation speed, so that by using the generated logistic regression model, it is possible to detect abnormalities in the rolling bearing 11 with even greater accuracy.

[0126] In the first to sixth embodiments, the detection device main body 21 may perform the process up to outputting the abnormality probability, and the user may determine whether or not there is an abnormality based on the output abnormality probability.

[0127] Seventh Embodiment Next, Fig. 16 is a block diagram showing the main components of a rolling bearing anomaly detection device according to a seventh embodiment. In the seventh embodiment, two or more types of logistic regression models are generated by the model generation unit 29. The two or more types of logistic regression models are then stored in the detection device main body 21. In this example, the six types of logistic regression models shown in the first to sixth embodiments are stored in the detection device main body 21.

[0128] The probability calculation unit 31 calculates two or more abnormality probabilities for the same rolling bearing 11 using two or more types of logistic regression models.

[0129] The determination unit 32 compares each of the multiple abnormality probabilities with a threshold value, for example 50%, and determines whether or not there is an abnormality in the rolling bearing 11 based on the number of abnormality probabilities that are equal to or greater than the threshold value.

[0130] For example, if the number of abnormality probabilities that are above the threshold is greater than or equal to the number of abnormality probabilities that were below the threshold, the judgment unit 32 judges that an abnormality exists, i.e., that damage has occurred to the rolling bearing 11.

[0131] Furthermore, if the number of abnormality probabilities that are equal to or greater than the threshold value is less than the number of abnormality probabilities that were less than the threshold value, the determination unit 32 determines that the rolling bearing 11 is in a normal state.

[0132] In this example, the probability calculation unit 31 has a first calculation unit 31a, a second calculation unit 31b, a third calculation unit 31c, a fourth calculation unit 31d, a fifth calculation unit 31e, and a sixth calculation unit 31f.

[0133] The first calculation unit 31a calculates the first abnormality probability using the same logistic regression model as in embodiment 1. The second calculation unit 31b calculates the second abnormality probability using the same logistic regression model as in embodiment 2. The third calculation unit 31c calculates the third abnormality probability using the same logistic regression model as in embodiment 3.

[0134] The fourth calculation unit 31d calculates the fourth abnormality probability using the same logistic regression model as in embodiment 4. The fifth calculation unit 31e calculates the fifth abnormality probability using the same logistic regression model as in embodiment 5. The sixth calculation unit 31f calculates the sixth abnormality probability using the same logistic regression model as in embodiment 6.

[0135] Other configurations and operations in the seventh embodiment are the same as those in the first to sixth embodiments.

[0136] In this type of abnormality detection device 20, two or more abnormality probabilities are calculated, and the presence or absence of an abnormality in the rolling bearing is determined based on the number of abnormality probabilities that is equal to or greater than a threshold value. This makes it possible to detect abnormalities in the rolling bearing 11 with even greater accuracy.

[0137] In the seventh embodiment, the number of types of logistic regression models is not limited to six, but may be five or less, or seven or more.

[0138] In the seventh embodiment, the criteria for determining whether or not an abnormality exists in the determination unit 32 are not limited to the above example, and for example, it may be determined that an abnormality exists when at least one abnormality probability is equal to or greater than a threshold value. Alternatively, it may be determined that an abnormality exists only when all abnormality probabilities are equal to or greater than a threshold value.

[0139] In addition, in the first to seventh embodiments, the generation of the logistic regression model may be performed by a model generation device separate from the anomaly detection device 20.

[0140] Furthermore, in embodiments 1 to 7, if a single piece of equipment, for example, an elevator hoist, is provided with multiple rolling bearings, the object of monitoring by abnormality detection device 20 need not be all of the rolling bearings, but may be only one of the rolling bearings.

[0141] Furthermore, in the first to seventh embodiments, the rolling bearing 11 is not limited to a rolling bearing for an elevator hoisting machine.

[0142] Each function of the detection device main body 21 in the first to seventh embodiments is realized by a processing circuit. Fig. 17 is a configuration diagram showing a first example of a processing circuit that realizes each function of the detection device main body 21 in the first to seventh embodiments. The processing circuit 100 in the first example is dedicated hardware.

[0143] The processing circuit 100 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof. Each function of the detection device main body 21 may be realized by a separate processing circuit 100, or all functions may be realized by the processing circuit 100.

[0144] 18 is a configuration diagram showing a second example of a processing circuit that realizes each function of the detection device main body 21 according to Embodiments 1 to 7. The processing circuit 200 of the second example includes a processor 201 and a memory 202.

[0145] In the processing circuit 200, each function of the detection device main body 21 is realized by software, firmware, or a combination of software and firmware. The software and firmware are written as programs and stored in the memory 202. The processor 201 realizes each function by reading and executing the programs stored in the memory 202.

[0146] The programs stored in memory 202 can be said to cause the computer to execute the procedures or methods of the above-mentioned components. Here, memory 202 refers to non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), and EEPROM (Electrically Erasable and Programmable Read Only Memory). Magnetic disks, flexible disks, optical disks, compact disks, minidisks, DVDs, and the like also fall under memory 202.

[0147] It should be noted that some of the functions of the above-described units may be realized by dedicated hardware, and other parts may be realized by software or firmware.

[0148] In this way, the processing circuit can realize the functions of each of the above-mentioned units by hardware, software, firmware, or a combination of these.

[0149] REFERENCE SIGNS LIST 11 Rolling bearing, 17 Acceleration sensor, 20 Abnormality detection device, 21 Detection device main body, 23 Vibration data acquisition unit, 24 Filter processing unit, 25 Overall value calculation unit, 26 Envelope analysis unit, 27 Peak height calculation unit.

Claims

1. A rolling bearing anomaly detection device comprising: a detection device main body which calculates an overall value and a peak height of a frequency in a frequency waveform caused by damage to the rolling bearing from vibration data obtained from an acceleration sensor provided in the rolling bearing, and calculates an abnormality probability which is the probability that damage has occurred inside the rolling bearing based on a plurality of reference values ​​including the overall value and the peak height, and a logistic regression model, wherein the logistic regression model is a logistic function generated by performing logistic regression analysis using a plurality of indices including the overall value and the peak height from learning data of a plurality of samples which includes normal data which is vibration data obtained by rotating a plurality of rolling bearings in a normal state, and abnormal data which is vibration data obtained by rotating a plurality of rolling bearings in a damaged state.

2. The rolling bearing abnormality detection device according to claim 1, wherein each of said plurality of indexes and said plurality of reference values ​​further includes a value obtained by multiplying said overall value by said peak height.

3. A rolling bearing abnormality detection device as claimed in claim 1 or 2, wherein each of said plurality of indexes and said plurality of reference values ​​further includes an operating time of said rolling bearing.

4. A rolling bearing abnormality detection device as claimed in claim 1 or 2, wherein each of said plurality of indicators and said plurality of reference values ​​further includes a total number of revolutions of said rolling bearing.

5. A rolling bearing abnormality detection device as described in any one of claims 1 to 4, wherein a threshold value for the abnormality probability is set in the detection device main body, and the detection device main body determines the presence or absence of an abnormality in the rolling bearing by comparing the calculated abnormality probability with the threshold value.

6. A rolling bearing abnormality detection device as described in claim 5, wherein the detection device main body calculates two or more abnormality probabilities for the same rolling bearing using two or more types of the logistic regression model, compares each of the two or more abnormality probabilities with the threshold value, and determines the presence or absence of an abnormality in the rolling bearing based on the number of abnormality probabilities that are equal to or greater than the threshold value.

7. A rolling bearing abnormality detection device as claimed in any one of claims 1 to 6, wherein the detection device main body comprises: a vibration data acquisition unit that acquires the vibration data from the acceleration sensor; a filter processing unit that applies filter processing to the vibration data acquired by the vibration data acquisition unit to generate noise-removed data, which is data from which noise has been removed; an overall value calculation unit that calculates the overall value from the noise-removed data; an envelope analysis unit that applies envelope processing to the noise-removed data and outputs a Fourier transformed frequency waveform; and a peak height calculation unit that calculates the peak height from the frequency waveform.

8. A method for generating a logistic regression model comprising the steps of: acquiring learning data of a plurality of samples including normal data, which is vibration data obtained by rotating a plurality of rolling bearings in a normal state, and abnormal data, which is vibration data obtained by rotating a plurality of rolling bearings in a damaged state; classifying the learning data into two values, the normal data and the abnormal data; and selecting a plurality of indices from the binary-classified learning data, performing a logistic regression analysis, and deriving, as a logistic regression model, a logistic function for calculating an abnormality probability, which is the probability that damage has occurred inside the rolling bearing.

9. A method for generating a logistic regression model as described in claim 8, wherein the multiple indexes include an overall value calculated by filtering the vibration data and a peak height of a frequency caused by damage to a rolling bearing, and the peak height is calculated from a frequency waveform obtained by performing envelope processing on the filtered data and carrying out frequency analysis.

10. The method for generating a logistic regression model according to claim 9, wherein the plurality of indices further includes a value obtained by multiplying the overall value by the peak height.

11. A method for generating a logistic regression model according to claim 9 or 10, wherein the plurality of indicators further includes an operating time of the rolling bearing.

12. A method for generating a logistic regression model according to claim 9 or 10, wherein the plurality of indicators further includes a total number of revolutions of the rolling bearing.

Images