Bearing defect diagnosis method

Abstract

The present invention relates to a bearing defect diagnosis method for diagnosing a defect of a bearing on the basis of the difference in frequency range and the difference in frequency amplitude of sound data between normal bearings and defective bearings. The bearing defect diagnosis method according to the present invention comprises: a step (a) of segmenting sound data collected by a microphone during a bearing operation process into predetermined time intervals; a step (b) of frequency-analyzing the sound data, which was segmented in step (a), to extract, as physical features, the difference in frequency range and frequency amplitudes of the sound data of normal bearings and defective bearings; and a step (c) of inputting the physical features, which were extracted in step (b), to an artificial neural network to diagnose whether a defect has occurred in a bearing.

Description

Method for diagnosing bearing defects

[0001] The present invention relates to a method for diagnosing a defect in a bearing, and more particularly, to a method for diagnosing a defect in a bearing based on a difference in a frequency range and a difference in frequency amplitude in sound data between a normal bearing and a defective bearing.

[0002] As is well known, a bearing is a machine element that limits relative motion for a desired movement and reduces friction between moving parts. It is a machine device (part) that supports a rotating or reciprocating shaft at a certain position and allows it to move freely.

[0003] Meanwhile, if a defect occurs in the bearing, such as a defect in its inner diameter, outer diameter, or ball, and grows, it is necessary to take measures such as checking for this early and replacing it, as this may cause damage to the device on which the bearing is mounted or cause various accidents.

[0004] However, it is difficult to diagnose bearing defects because they generate acoustic signals of different frequency bands depending on the location (inner diameter, outer diameter, ball, etc.) within the bearing where the defect occurred and the size of the defect.

[0005] In diagnosing defects in conventional bearings, Fourier transform is performed on sound data collected while the bearing is operating using a non-contact sensor such as a microphone to extract frequency components, a spectrogram image is created using the extracted frequency components, and this is classified using a machine learning algorithm to diagnose the presence and cause of the failure.

[0006] Here, a spectrogram is a tool for visualizing and understanding sound or waves, and is a graph that combines the characteristics of a waveform and a spectrum. While a typical waveform graph shows changes in amplitude over time, and a spectrum graph shows changes in amplitude over frequency, a spectrogram displays the differences in amplitude according to changes in the time and frequency axes using different print densities or display colors.

[0007] However, the Fourier transform has a limitation in that it cannot simultaneously improve frequency resolution and time resolution because it decomposes the signal using infinitely oscillating sine and cosine functions as basis functions under the assumption that the signal does not change in time. Accordingly, there was a problem in that accuracy was reduced when performing frequency analysis over a wide range of frequency domains.

[0008] In addition, there was a problem that it was difficult to accurately diagnose bearing defects by transmitting inaccurate frequency analysis results to the machine learning algorithm when information related to bearing defects was less distributed within the generated spectrogram image.

[0009] The main task of the present invention is to provide a bearing defect diagnosis method that diagnoses a bearing defect based on the difference in the frequency range and the difference in the frequency amplitude of sound data between a normal bearing and a defective bearing.

[0010] Another object of the present invention is to provide a bearing defect diagnosis method that can more accurately diagnose a bearing defect by effectively visualizing physical characteristics appearing in sound data as a defect occurs in the bearing in the form of adaptive scalar image data.

[0011] In order to achieve the above object, a bearing defect diagnosis method according to the present invention comprises the steps of: (a) dividing sound data collected by a microphone during a bearing operation process into a predetermined time size; (b) analyzing the sound data divided in step (a) for frequency to extract differences in frequency ranges having large frequency amplitude values ​​and relative frequency amplitude differences in the sound data of a normal bearing and a defective bearing as physical characteristics; and (c) inputting the physical characteristics extracted in step (b) into an artificial neural network to diagnose whether a bearing defect has occurred.

[0012] In the above-described configuration, the frequency range having a large value of frequency amplitude in the normal bearing is 0 to 100 Hz, which is the shaft rotation frequency range, and the frequency range having a large value of frequency amplitude in the defective bearing is 200 to 672 Hz, which is the defective frequency range.

[0013] The frequency amplitude in the shaft rotation frequency region remains uniform regardless of the occurrence of a defect, but the frequency amplitude in the defect frequency region increases as the size of the defect increases.

[0014] (b) The physical characteristics in step (b) are visualized in the form of an image through a visualization process using a scalogram.

[0015] In the visualization process through the scalogram, the sound data is wavelet transformed to extract the frequency amplitude, and then the maximum value of the color mapping is set to the 99th percentile (A) of all frequency amplitudes so that the brightest color is secured at least 1% when color mapping the scalogram. 1% ) is designated.

[0016] 0~99th percentile (A 1% ) The frequency amplitude of the range is divided into five equal sections and assigned different colors.

[0017] 99th percentile (A 1% ) greater than the frequency amplitude range Merge into sections.

[0018] The artificial neural network is a pre-trained CNN model based on the physical characteristics of defect occurrence verified through experiments.

[0019] According to the bearing defect diagnosis method of the present invention, since the length of sound data required to diagnose a bearing defect in the data measurement step is sufficient to be at least 0.25 seconds, a large-capacity data storage device is not required.

[0020] Additionally, since the microphone, which is a sensor that measures sound data, is a non-adhesive sensor, multiple bearings can be monitored with a single device when moving the microphone and sequentially monitoring multiple rotating devices.

[0021] In addition, since adaptive scalograms can be developed and extracted based on physical characteristics of defect occurrence verified through actual experiments, sufficient physical basis for defect diagnosis results can be secured.

[0022] By pre-training the CNN (Convolutional Neural Network) model used as a classifier in the defect diagnosis stage with adaptive scalograms of defect states and normal states generated through experiments, it has the effect of being able to independently derive optimal criteria for defect diagnosis based on big data.

[0023] Figure 1 is a diagram showing the sound data measured from a normal bearing and a defective bearing and the frequency components of the normal bearing and the defective bearing extracted through frequency analysis.

[0024] Figure 2 is a flowchart for explaining a method for diagnosing a bearing defect according to one embodiment of the present invention.

[0025] Figure 3 is a drawing of the positions of the non-contact microphone and bearing installed to measure sound data in an experimental device simulating a rotating machine.

[0026] Figure 4 is a diagram for explaining a method for generating a general scalogram and an adaptive scalogram using frequency components extracted through frequency analysis.

[0027] FIG. 5 is a diagram showing the structure and mechanism of a CNN used as a classifier in the defect diagnosis step of the bearing defect diagnosis method of the present invention.

[0028] Hereinafter, a preferred embodiment of a bearing defect diagnosis method of the present invention will be described in detail with reference to the attached drawings. Detailed descriptions of well-known functions and configurations that may unnecessarily obscure the gist of the invention will be omitted. Furthermore, since the embodiments of the present invention are provided to more fully explain the present invention to a person of average skill in the art, the shapes and sizes of elements in the drawings may be exaggerated for clearer explanation.

[0029] In the bearing defect diagnosis method of the present invention, a bearing defect is diagnosed based on the difference in the frequency range and the frequency amplitude of sound data between a normal bearing and a defective bearing.

[0030] More specifically, we propose a method to improve diagnostic accuracy by extracting frequency components of sound data collected while the bearing is operating using wavelet transform, which can improve both frequency resolution and time resolution instead of the conventional Fourier transform method, visualizing these frequency components in the form of an adaptive scalogram, which is image data, and thereby delivering accurate frequency analysis results related to bearing defects to a CNN model, which is a deep learning algorithm effective for image classification.

[0031] Figure 1 is a diagram showing the sound data measured from a normal bearing and a defective bearing, respectively, and the frequency components of a normal bearing and a defective bearing extracted through frequency analysis.

[0032] First, it can be seen that the frequency component of the sound data of Fig. 1 (a) measured from a normal bearing using a microphone, which is a non-contact sensor, exists only in the shaft rotation frequency range of 0 to 100 Hz due to shaft rotation, as can be seen in Fig. 1 (b).

[0033] On the other hand, the frequency components of the sound data of (c) of Fig. 1 measured in a bearing having a defect in the inner diameter, outer diameter, or ball constituting the bearing are mixed in the defect frequency range of 200 to 672 Hz, which occurs when the defect comes into contact with another part, and the shaft rotation frequency range of 0 to 100 Hz, as can be seen in (d) of Fig. 1.

[0034] In Fig. 1, BPFO (Ball Pass Frequency Outer race) represents a frequency component due to a defect in the outer diameter of the bearing, BPFI (Ball Pass Frequency Inner race) represents a frequency component due to a defect in the inner diameter of the bearing, and BSF (Ball Spin Frequency) represents a frequency component due to a defect in the ball of the bearing.

[0035] In particular, it can be seen that the frequency amplitude of the shaft rotation frequency range of 1 to 100 Hz (Zone 1) in the sound data remains constant regardless of the occurrence of a bearing defect, but the frequency amplitude of the defect frequency range of 200 to 672 Hz (Zone 2) gradually increases as the size of the defect increases.

[0036] As a result, when a defect occurs in the bearing, the frequency amplitude in the defect frequency region becomes larger than the frequency amplitude in the shaft rotation frequency region, and as the defect grows, the difference in relative frequency amplitude between the two frequency regions gradually increases.

[0037] In the defect diagnosis method of the present invention, the occurrence of a defect in a bearing and its cause are diagnosed based on the difference in each frequency range and its frequency amplitude, which are physical characteristics between the normal bearing and the defective bearing described above.

[0038] FIG. 2 is a flowchart illustrating a bearing defect diagnosis method according to an embodiment of the present invention. As illustrated in FIG. 2, according to the bearing defect diagnosis method according to an embodiment of the present invention, first, in step S10, a microphone, which is a non-contact sensor, is used to collect sound data around the bearing during the process of operating the bearing. Since sound data of at least 0.25 seconds, and preferably 1 to 10 seconds, is sufficient for diagnosing a bearing defect, a large-capacity data storage device is not required. Next, the sound data collected in this way is divided into units of 0.25 seconds, which is the minimum length required for diagnosing a bearing defect, and stored in a data storage device.

[0039] Figure 3 is a diagram illustrating the locations of a non-contact microphone and bearing installed in an experimental device simulating a rotating machine to measure sound data. As illustrated in Figure 3, during the sound data collection process, the non-contact microphone is positioned at a distance of, for example, 100 mm from the bearing being diagnosed, and sound data generated during the bearing's operation is collected.

[0040] In this way, since the defect diagnosis method of the present invention uses a microphone, which is a non-contact sensor, to collect sound data, it is possible to sequentially monitor multiple rotating machines by moving one microphone.

[0041] Returning to FIG. 2 again, in step S20, the 0.25 second long sound data collected in step S10 is subjected to wavelet transformation to extract frequency components, and only the frequency components of 1 to 672 Hz, which include both the shaft rotation frequency range and the bearing defect frequency range, are extracted.

[0042] The wavelet transform used in the defect diagnosis method of the present invention is a mathematical tool for overcoming the limitations of the Fourier transform, and can be useful for analyzing data whose features change at different scales, for example, data whose features change, such as frequency changing over time in the case of a signal, or a trend that changes transiently or slowly.

[0043] Specifically, the Fourier transform decomposes a signal into sine waves of specific frequencies, while the wavelet transform decomposes the signal into wavelets, each shifted and scaled. Unlike sine waves, wavelets are rapidly decaying oscillations, similar to waves, allowing them to represent data across multiple scales.

[0044] Returning to FIG. 2 again, in step S30, a scalogram in the form of image data is generated to relatively compare the frequency amplitude of the shaft rotation frequency range of 1 to 100 Hz among the extracted frequency components with the frequency amplitude of the bearing defect frequency range of 200 Hz or more. However, instead of a general scalogram, an adaptive scalogram specifically proposed in the present invention is generated to visualize the physical characteristics according to the bearing defect.

[0045] Here, a scalogram is a method of signal visualization that uses the wavelet transform instead of the Fourier transform. Unlike a spectrogram, a scalogram doesn't display waveform information because its horizontal axis is time and its vertical axis is frequency. Instead, it allows for a visual understanding of temporal changes across frequency bands.

[0046] Figure 4 is a diagram illustrating a method for generating a general scalogram and an adaptive scalogram using frequency components extracted through frequency analysis. As illustrated in Figure 4 (a), a general scalogram visualizes frequency amplitude by normalizing frequency amplitude values ​​to, for example, values ​​between 0 and 255 in a minimum-maximum manner and displaying them in different colors. As the frequency amplitude increases, it is displayed in a brighter color (yellow), and as the frequency amplitude decreases, it is displayed in a gradually darker color (blue).

[0047] However, when using a general scalogram, the area that appears in the brightest color (yellow) accounts for less than 0.2% of the entire image, so the CNN model, a deep learning algorithm that is effective in image classification, could not properly capture the physical characteristics of defect occurrence verified through actual experiments, resulting in a defect diagnosis accuracy of only 94.1%.

[0048] In order to improve such low defect diagnosis accuracy, the defect diagnosis method of the present invention generates an adaptive scalogram. First, in step S31, the 99th percentile (A) is calculated for all frequency amplitudes extracted from 1 to 672 Hz so that the area that appears in the brightest color during color mapping of the scalogram is secured to be at least 1% of the entire image area. 1% ) is extracted, the brightest color is assigned to all frequency areas with frequency amplitudes greater than that value.

[0049] Next, in step S32, a value less than the 99th percentile (0 to A) is calculated as shown in (b) of Fig. 4. 1% ) is divided into 5 parts according to size, and one color is assigned to each section, and in step S33, A 1% Larger frequency amplitude range By merging into sections, it can be seen that in the adaptive scalogram proposed in the present invention, the area that appears in the brightest color (yellow) occupies 2.5% of the entire image, which is significantly higher than that of a general scalogram.

[0050] In step S34, a color bar is constructed by assigning different colors to each of the five sections thus generated, and in step S35, an adaptive scalogram in the form of image data is generated with the colors of the color bar thus constructed. Accordingly, as can be seen in the lower part of Fig. 4 (b), it can be seen that the physical characteristics in the 200-672 Hz region, which is the bearing defect frequency region, are visualized more clearly than the general scalogram shown in Fig. 4 (a) when the adaptive scalogram of the present invention is used. In addition, the influence of outliers such as very large frequency amplitudes and noise appearing in the low frequency region on defect diagnosis can be reduced.

[0051] Returning to FIG. 2 again, in step S40, the adaptive scalogram generated in step S30 is classified using a pre-learned CNN model based on physical characteristics according to the occurrence of a defect verified through an actual experiment to diagnose a defect in the bearing. FIG. 5 is a diagram showing the structure and mechanism of a CNN used as a classifier in the defect diagnosis step of the defect diagnosis method of the present invention.

[0052] As described above, in the bearing defect diagnosis method of the present invention, the physical characteristics of the sound signal data according to the presence or absence of a defect in the bearing are clearly visualized through a scalogram, and the 99th percentile (A) among the frequency amplitudes 1% ) is normalized to the maximum value, and the 99th percentile (0 to A) is reduced to a significant extent so that the total number of colors is reduced to 5. 1%) is divided into five equal parts and a color is assigned to each section. As a result, when the adaptive scalogram is used, the fault diagnosis accuracy is improved by 5.7% compared to when the general scalogram is used, and the natural fault of the bearing is diagnosed with an accuracy of 99.8%, which is close to 100%.

[0053] The bearing defect diagnosis method of the present invention can be applied to all fields in which bearings are used, such as production and transportation fields, and can diagnose defects in bearings operating inside rotating equipment with very high accuracy.

[0054] While the present invention has been described with reference to one embodiment illustrated in the accompanying drawings, this is merely exemplary. Those skilled in the art will appreciate that various modifications and equivalent embodiments are possible. For example, the 99th percentile or the number of color bars may be appropriately increased or decreased.

[0055] Accordingly, the true scope of protection of the present invention should be determined solely by the appended claims.

[0056] According to the bearing defect diagnosis method of the present invention, since the length of sound data required to diagnose a bearing defect in the data measurement step is sufficient to be at least 0.25 seconds, a large-capacity data storage device is not required.

[0057] Additionally, since the microphone, which is a sensor that measures sound data, is a non-adhesive sensor, multiple bearings can be monitored with a single device when moving the microphone and sequentially monitoring multiple rotating devices.

[0058] In addition, since adaptive scalograms can be developed and extracted based on physical characteristics of defect occurrence verified through actual experiments, sufficient physical basis for defect diagnosis results can be secured.

[0059] By pre-training the CNN (Convolutional Neural Network) model used as a classifier in the defect diagnosis stage with adaptive scalograms of defect states and normal states generated through experiments, it has the effect of being able to independently derive optimal criteria for defect diagnosis based on big data.

Claims

1. Step (a) of dividing sound data collected by a microphone during bearing operation into set time sizes; (a) Step (b) of analyzing the frequency of the sound data divided in step (a) to extract the difference in the frequency domain and the frequency amplitude of the sound data of the normal bearing and the defective bearing as physical characteristics; and A method for diagnosing a bearing defect, comprising a step (c) of inputting the physical features extracted in step (b) into an artificial neural network to diagnose whether a defect has occurred in the bearing.

2. In claim 1, A method for diagnosing a bearing defect, characterized in that the frequency range of a normal bearing is 0 to 100 Hz, which is a shaft rotation frequency range, and the frequency range of a defective bearing is 1 to 100 Hz, which is a shaft rotation frequency range, and 200 to 672 Hz, which is a defective frequency range.

3. In claim 2, A method for diagnosing a bearing defect, characterized in that the frequency amplitude in the defect frequency range increases as the size of the defect increases.

4. In claim 3, (b) A method for diagnosing a bearing defect, characterized in that the physical characteristics in step (b) are visualized in the form of an image through a visualization process using a scalogram.

5. In claim 4, In the visualization process through the scalogram, the sound data is subjected to wavelet transformation to extract the frequency amplitude, and then the maximum value of the color mapping is set to the 99th percentile (A) of all frequency amplitudes so that the brightest color is secured at least 1% when color mapping the scalogram. 1% ) is characterized by designating a bearing defect.

6. In claim 5, 0~99th percentile(A 1% ) A bearing defect diagnosis method characterized by dividing the frequency amplitude of a range into five equal sections and assigning different colors to each section.

7. In claim 6, 99th percentile (A 1% ) greater than the frequency amplitude range A method for diagnosing a bearing defect, characterized by merging into sections.

8. In any one of claims 1 to 7, A bearing defect diagnosis method characterized by an artificial neural network being a CNN model that has been pre-learned based on the physical characteristics of defect occurrence verified through experiments.

Images