A CNN Rolling Bearing Fault Diagnosis Method Based on Secondary Signal Enhancement

Abstract

The present invention relates to the technical field of bearing fault diagnosis, and particularly relates to a CNN rolling bearing fault diagnosis method based on secondary signal enhancement, including modeling using a public dataset to obtain a model; designing a trapezoidal filter; based on the trapezoidal filter and a band-pass filter, enhancing the signals of the public dataset according to known fault types to obtain training sample images; using the training sample images to train the model to obtain an identification model; based on the trapezoidal filter and the band-pass filter, enhancing the signals of actual samples according to known fault types to obtain actual sample images; putting the actual sample images into the identification model for fault diagnosis to obtain a diagnosis result. The present invention performs signal enhancement twice, enhancing the signals through the public dataset of the filter, thereby solving the problem that weak signals cannot be used for fault detection, resulting in a reduction in the accuracy of bearing detection.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and particularly to a CNN rolling bearing fault diagnosis method based on secondary signal enhancement. Background Art

[0002] Fault diagnosis of rolling bearings generally includes knowledge-based, model-based, and data-driven diagnosis methods. In recent years, data-driven methods have been greatly studied and applied, among which convolutional neural networks based on deep learning have been more fully applied.

[0003] Utilizing the CNN's ability to mine non-linear features, CNN diagnosis methods based on time-frequency signals have been fully studied.

[0004] Although rich research results have been achieved in the fault diagnosis of rolling bearings based on the deep convolutional network CNN, in terms of how to construct better feature images as the input of CNN, it is unable to detect faults in weak signals, resulting in a reduction in the accuracy of bearing detection. Summary of the Invention

[0005] The purpose of the present invention is to provide a CNN rolling bearing fault diagnosis method based on secondary signal enhancement, aiming to solve the problem that faults in weak signals cannot be detected, resulting in a reduction in the accuracy of bearing detection.

[0006] To achieve the above purpose, the present invention provides a CNN rolling bearing fault diagnosis method based on secondary signal enhancement, including the following steps:

[0007] Use a public dataset for modeling to obtain a model;

[0008] Design a trapezoidal filter and a band-pass filter;

[0009] Based on the trapezoidal filter and the band-pass filter, enhance the signals of the public dataset according to known fault types to obtain training sample images;

[0010] Use the training sample images to train the model to obtain an identification model;

[0011] Based on the trapezoidal filter and the band-pass filter, enhance the signals of actual samples according to known fault types to obtain actual sample images;

[0012] Put the actual sample images into the identification model for fault diagnosis to obtain a diagnosis result.

[0013] Wherein, the public dataset includes rotational speed frequency, sampling frequency, and bearing structure parameters.

[0014] Among them, the specific method of using the public data set for modeling to obtain the model is as follows:

[0015] Modeling is performed through the rotational speed frequency, the sampling frequency, and the bearing structure parameters to obtain the model.

[0016] Among them, the specific method of strengthening the signals of the public data set according to the known fault types based on the trapezoidal filter and the band-pass filter to obtain the training sample images is as follows:

[0017] Select samples of each fault according to the known fault types;

[0018] Make an amplitude spectrum based on the parameters of the samples to obtain the training amplitude spectrum;

[0019] Multiply the training amplitude spectrum by the trapezoidal filter to obtain the training time-domain filtered signal;

[0020] Perform short-time Fourier transform on the training time-domain filtered signal and then adjust the size to obtain the training grayscale image;

[0021] Use the band-pass filter to enhance the texture of the training grayscale image to obtain the training sample image.

[0022] Among them, the specific method of strengthening the signals of the actual samples according to the known fault types based on the trapezoidal filter and the band-pass filter to obtain the actual sample images is as follows:

[0023] Make an amplitude spectrum based on the actual samples and obtain the actual amplitude spectrum;

[0024] Multiply the actual amplitude by the trapezoidal filter to obtain the actual time-domain filtered signal;

[0025] Perform short-time Fourier transform on the actual time-domain filtered signal and then adjust the size to obtain the actual grayscale image;

[0026] Use the band-pass filter to enhance the texture of the actual grayscale image to obtain the actual sample image.

[0027] A CNN rolling bearing fault diagnosis method based on secondary signal enhancement of the present invention uses a public dataset for modeling to obtain a model; designs a trapezoidal filter; based on the trapezoidal filter and the band-pass filter, signals of the public dataset are enhanced according to known fault types to obtain training sample images; the model is trained using the training sample images to obtain an identification model; based on the trapezoidal filter and the band-pass filter, signals of actual samples are enhanced according to known fault types to obtain actual sample images; the actual sample images are put into the identification model for fault diagnosis to obtain a diagnosis result. The present invention performs signal enhancement twice on the public dataset and the actual samples respectively through the trapezoidal filter and the band-pass filter, and performs signal enhancement on the public data through the filter, thereby solving the problem that weak signals cannot be used for fault detection, resulting in a reduction in the accuracy of bearing detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0029] Figure 1 is a flowchart of a CNN rolling bearing fault diagnosis method based on secondary signal enhancement provided by the present invention.

[0030] Figure 2 is a flowchart of enhancing signals of the public dataset according to known fault types based on the trapezoidal filter and the band-pass filter to obtain training sample images.

[0031] Figure 3 is a flowchart of enhancing signals of actual samples according to known fault types based on the trapezoidal filter and the band-pass filter to obtain actual sample images.

[0032] Figure 4 is a schematic diagram of a trapezoidal oscilloscope. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] The following will describe in detail the embodiments of the present invention. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present invention and should not be construed as a limitation of the present invention.

[0034] Please refer to Figures 1 to 4, the present invention provides a CNN rolling bearing fault diagnosis method based on secondary signal enhancement,

[0035] S1 Use the public dataset for modeling to obtain a model;

[0036] Specifically, the public dataset includes rotational speed frequency, sampling frequency, and bearing structure parameters.

[0037] Modeling is performed through the rotational speed frequency, the sampling frequency, and the bearing structure parameters to obtain the model.

[0038] S2 Design a trapezoidal filter and a band-pass filter;

[0039] Specifically, the trapezoidal filter is for the time-frequency image to perform targeted filtering on the typical fault characteristic frequency using a band-pass (Gabor) filter to form secondary signal enhancement. The trapezoidal pass filter has 4 fault characteristic frequencies f T , f O , f B , f I The smallest frequency f min and the largest frequency f max ; the long side of the trapezoidal filter is f1 f2, f1 = f min -0.3*f T , f2 = f max +0.3*f T ; the amplitudes of the 4 characteristic frequencies are h T , h O , h B , h I , and its largest amplitude is h max , thus the long side and height of the trapezoidal filtering have been determined.

[0040] S3 Based on the trapezoidal filter and the band-pass filter, enhance the signals of the public dataset according to the known fault types to obtain training sample images;

[0041] S31 Select samples of each fault according to the known fault types;

[0042] Specifically, the known fault types are inner race fault, outer race fault, rolling element fault, and cage fault, which are respectively: f T , f O , f B , f I ;

[0043] Inner race fault frequency:

[0044] f I =n×f BPI ;

[0045] Outer race fault frequency:

[0046] f O = n × f BPO ;

[0047] Rolling element fault frequency:

[0048] f B = n × f BS ;

[0049] Cage outer race fault frequency:

[0050] f T = n × f FT ;

[0051] Where f BPI , f BPO , f BS , f FT are the spin frequencies of the inner race, outer race, cage and rolling elements of the bearing respectively, and can be calculated according to the bearing model.

[0052] S32 Make an amplitude spectrum based on the parameters of the sample to obtain a training amplitude spectrum;

[0053] S33 Multiply the training amplitude spectrum by the trapezoidal filter to obtain a training time-domain filtered signal;

[0054] Specifically, multiply the training amplitude spectrum of the collected time-domain signal by the trapezoidal filter, and perform an inverse Fourier transform on the obtained spectrum to obtain the training time-domain filtered signal.

[0055] S34 Perform short-time Fourier transform on the training time-domain filtered signal and then adjust the size to obtain a training grayscale image;

[0056] Specifically, perform short-time Fourier transform STFT on the filtered time-domain signal, take the time domain length as 1024, remove the DC component; the window length is 64, the window type is Hanning window, and the overlap rate is 50%; take the positive frequency part of the amplitude spectrum of STFT; assume the discrete time range of this part is (0, t max ), t max is the time length, with the unit of points; the normalized frequency range is (0, f max ), f max is the frequency length, with the unit of points; the positive frequency amplitude spectrum value range is (h min , h max ), and convert this two-dimensional signal to the sample standard size:

[0057] C1: Convert the amplitude of each point in the positive frequency STFT to grayscale

[0058]

[0059] I (x,y) is the grayscale value converted to 0 - 255;

[0060] h (x,y) is the amplitude spectrum value before conversion; x and y are the time coordinate points and frequency coordinate points of the positive frequency amplitude spectrum respectively; for the training grayscale image that has been converted, through stretching or compression, it is unified to 256 * 256.

[0061] S35 Use the band - pass filter to perform texture enhancement on the training grayscale image to obtain a training sample image.

[0062] Specifically, the designed Gabor filter needs to further strengthen 4 characteristic frequencies to enhance the texture. Let f T , f O , f B , f I After sorting from small to large as f1, f2, f3, f4, the center frequency bands of the designed Gabor filter are f1, f2, f3, f4 respectively. Considering the angle, the angle is designed in 8 directions. However, since the texture of the image after short - time Fourier transform accounts for a larger proportion in the horizontal direction, the 8 designed directions do not consider the vertical direction. The angles are: 0°, 30°, 45°, 60°, 90°, 120°, 135°, 150°. The time - window scales are taken respectively In this way, there are a total of 24 Gabor filters with a template size of 32 * 32. Filter the 256 * 256 sample image through the Gabor filter to form time - domain image enhancement.

[0063] S4 Use the training sample image to train the model to obtain an identification model;

[0064] S5 Based on the trapezoidal filter and the band - pass filter, perform signal enhancement on the actual sample according to the known fault type to obtain an actual sample image;

[0065] S51 Select the samples of each fault according to the known fault type;

[0066] S52 Multiply the actual amplitude by the trapezoidal filter to obtain an actual time - domain filtered signal;

[0067] S53 Perform short - time Fourier transform on the actual time - domain filtered signal and then adjust the size to obtain an actual grayscale image;

[0068] S54 Use the band - pass filter to perform texture enhancement on the actual grayscale image to obtain an actual sample image;

[0069] S6 Put the actual sample image into the identification model for fault diagnosis to obtain a diagnosis result.

[0070] The above-disclosed is only a preferred embodiment of a CNN rolling bearing fault diagnosis method based on secondary signal enhancement of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. A CNN rolling bearing fault diagnosis method based on secondary signal enhancement, characterized in that, It includes the following steps: Use a publicly available dataset for modeling to obtain a model; Design a trapezoidal filter and a band - pass filter; Based on the trapezoidal filter and the band - pass filter, strengthen the signals of the publicly available dataset according to known fault types to obtain training sample images; Use the training sample images to train the model to obtain an identification model; Based on the trapezoidal filter and the band - pass filter, strengthen the signals of actual samples according to known fault types to obtain actual sample images; Put the actual sample images into the identification model for fault diagnosis to obtain a diagnosis result; The specific method of strengthening the signals of the publicly available dataset according to known fault types based on the trapezoidal filter and the band - pass filter to obtain training sample images is as follows: Select samples of each fault according to known fault types; Make an amplitude spectrum based on the parameters of the samples to obtain a training amplitude spectrum; Multiply the training amplitude spectrum by the trapezoidal filter to obtain a training time - domain filtered signal; Perform short - time Fourier transform on the training time - domain filtered signal and then adjust the size to obtain a training grayscale image; Use the band - pass filter to enhance the texture of the training grayscale image to obtain training sample images.

2. A CNN rolling bearing fault diagnosis method based on secondary signal strengthening according to claim 1, characterized in that The publicly available dataset includes rotational speed frequency, sampling frequency, and bearing structure parameters.

3. A CNN rolling bearing fault diagnosis method based on secondary signal strengthening according to claim 2, characterized in that The specific method of using the publicly available dataset for modeling to obtain a model is as follows: Perform modeling through the rotational speed frequency, the sampling frequency, and the bearing structure parameters to obtain the model.

4. A CNN rolling bearing fault diagnosis method based on secondary signal strengthening according to claim 1, characterized in that The specific method of strengthening the signals of actual samples according to known fault types based on the trapezoidal filter and the band - pass filter to obtain actual sample images is as follows: Make an amplitude spectrum based on the actual samples and obtain an actual amplitude spectrum; Multiply the actual amplitude spectrum by the trapezoidal filter to obtain an actual time - domain filtered signal; Perform short - time Fourier transform on the actual time - domain filtered signal and then adjust the size to obtain an actual grayscale image; Use the band - pass filter to enhance the texture of the actual grayscale image to obtain actual sample images.

Images