Rolling bearing fault classification method based on variable scale energy decomposition and SR

By adopting variable-scale energy decomposition and signal noise reduction technology in rolling bearing fault diagnosis, combined with the method of the pole learning machine, the problem of inaccurate signal decomposition in the existing technology is solved, and the accuracy and recognition ability of fault classification are significantly improved.

CN119089309APending Publication Date: 2024-12-06HUNAN UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411252389.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-12-06

Smart Images

  • Figure CN119089309A_ABST
    Figure CN119089309A_ABST
Patent Text Reader

Abstract

The invention relates to a rolling bearing fault classification method based on variable scale energy decomposition and stochastic resonance (SR), belongs to the field of fault classification and signal processing, and particularly relates to a rolling bearing fault classification method based on the combination of variable scale energy decomposition (VSED) and stochastic resonance (SR) to form V-S System. The method comprises the following steps: firstly, decomposing a rolling bearing vibration signal for multiple times by using VSED, and selecting a signal with a relatively large fault frequency amplitude as an input signal of an SR system by calculating the frequency of each component; then, the component is processed by utilizing SR (Scheduling Register); and finally, classifying the processed signals by using an extreme learning machine (ELM) to realize state recognition of the rolling bearing signals. The invention provides a brand-new signal decomposition method which is applied to rolling bearing fault classification, and effective fault state identification can be carried out. The method has the advantages of multiple adjustable parameters, simple principle and high fault classification accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a rolling bearing fault classification method based on variable scale energy decomposition and SR, and belongs to the field of mechanical fault classification and signal processing. Background Art

[0002] In industrial production activities, in order to maintain the normal operation of mechanical equipment, it is extremely important to diagnose its faults in a timely manner. Once a mechanical equipment fails, the economic losses caused are immeasurable, and may even cause casualties. Therefore, in order to prevent the occurrence of failures in a timely manner, it is particularly important to find a high-accuracy fault diagnosis method. As one of the most common components in modern mechanical equipment, rolling bearings are widely used in many fields such as metallurgy, chemical industry, and machinery manufacturing. Most mechanical equipment will vibrate during operation, and different states of the equipment will produce different vibrations. The operating state of the machine can be identified through the collected vibration data. Therefore, feature extraction based on vibration data and in-depth mining of fault characteristics are one of the focuses of rolling bearing fault diagnosis.

[0003] Signal decomposition is one of the most widely used data processing methods. It is often used to process rolling bearing vibration signals and decompose rolling bearing vibration signals into multiple different components, thereby filtering out signals that are irrelevant to fault characteristic signals and obtaining signals that contain more fault characteristics, thereby achieving the classification task of rolling bearing fault vibration signals. However, most of the existing signal decomposition algorithms decompose signals into multiple components in a single fixed mode, which makes it difficult to achieve accurate decomposition of signals and fail to meet more satisfactory requirements. Therefore, it is difficult to improve the accuracy of rolling bearing fault classification.

[0004] Therefore, finding a fault signal processing method to improve the accuracy of rolling bearing fault classification is one of the key issues to be improved in this field. Summary of the invention

[0005] In view of the above problems, the present invention provides a rolling bearing fault classification method based on variable scale energy decomposition and SR to improve the accuracy of fault classification. The rolling bearing vibration signal is decomposed multiple times using the VSED algorithm, and the frequency of each short-term trend component is calculated, and the component with a large fault frequency amplitude is selected; then the component is denoised using SR; finally, the data set is divided, and different fault data are identified and classified using ELM.

[0006] The technical solution adopted by the present invention is: a rolling bearing fault classification method based on variable scale energy decomposition and SR, and the specific steps are as follows:

[0007] A1: First, the VSED algorithm is used to decompose the original vibration signal of the rolling bearing multiple times to obtain several short-term trend components.

[0008] A2: Calculate the frequency information of several short-term trend components and select the component with larger fault frequency amplitude.

[0009] A3: Use SR to perform noise reduction processing on the components obtained in step A2.

[0010] A4: The signal obtained in step A3 is randomly divided into a data set, and the ELM is used as a classifier to identify the rolling bearing state.

[0011] Specifically, the VSED algorithm principle in step A1 is as follows:

[0012] Given a signal to be decomposed , define a signal extraction operator , which is used to extract low-frequency information from the original signal to be decomposed and retain it in the long-term trend component. The remaining component is the short-term trend component containing high-frequency information, as shown in the following formula:

[0013]

[0014] in, is the long-term trend component, is the remaining short-term trend component.

[0015] Assumptions For a time series real-valued signal, use Represents a real-valued signal For convenience, we define In the interval Definition and , and in the interval superior Available, so define a signal extraction operator , in the continuous local extreme point interval The Gaussian filter template is generated using the probability density function of the normal distribution , the density function of the normal distribution is as follows:

[0016]

[0017] The baseline signal was extracted according to the following formula:

[0018]

[0019] is the Gaussian filter function, according to , , get the Gaussian filter template with a window size of 3 Filtering is performed. ,make ,and .available ,Right now .

[0020] Secondly .and Calculated according to the following formula:

[0021]

[0022] in, , usually taken .and Calculated according to the following formula:

[0023]

[0024] in, , usually taken .

[0025] After a decomposition, a long-term trend component is obtained. Defining the short-term trend component , as shown in the following formula:.

[0026]

[0027] This decomposition algorithm can extract signals according to different frequencies. The long-term trend component often contains low-frequency information, while the high-frequency information is retained in the short-term trend component. If you want to get multiple components, you can perform multiple decompositions.

[0028] Specifically, the SR system used in the SR noise reduction process in step A3 is:

[0029]

[0030] Among them, the parameters , is the input signal of the above SR system, is the output signal of the above SR system.

[0031] Specifically, in step A4, the output signal of SR As input to ELM.

[0032] Among them, the principle of ELM is as follows:

[0033] ELM is a typical single hidden layer feedforward neural network. Its network structure consists of input layer, hidden layer and output layer. Assume that the number of neurons in the hidden layer is The input layer has neurons, and the output layer has neurons, assuming The training set input matrix of samples And the output matrix They are:

[0034]

[0035] Assume that the hidden layer neuron activation function is , the network output matrix for:

[0036]

[0037] Among them, ω represents the connection weight between the input layer and the hidden layer, is the neuron threshold of the hidden layer,

[0038] The above formula can be expressed as:

[0039] in, The output matrix The transpose of is the hidden layer output matrix, and its specific form is:

[0040]

[0041] When the activation function When infinitely differentiable, ω and It can be randomly determined before network training and remains unchanged during network training. It can be obtained by solving the least squares solution of the following system of equations;.

[0042]

[0043] Its solution is

[0044] The beneficial effects of the present invention are:

[0045] (1) The VSED algorithm proposed in the present invention adjusts the signal decomposition effect by changing parameters during the signal decomposition process, and has the advantages of many adjustable parameters and good signal decomposition effect.

[0046] (2) The SR system introduced in the present invention can perform noise absorption processing on the signal obtained by the VSED algorithm, amplify weak fault signals, and highlight fault characteristics.

[0047] (3) The ELM classifier introduced in the present invention can obtain the optimal solution after only one calculation and has the advantages of extremely fast running speed and excellent generalization performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the VSED decomposition baseline signal diagram;

[0049] Figure 2 is a flow chart of the method of the present invention;

[0050] Figure 3 It is a signal chart of multiple short-term trend components;

[0051] Figure 4 It is the test set classification result diagram; DETAILED DESCRIPTION

[0052] The present invention is further described below in conjunction with the accompanying drawings and specific implementation examples.

[0053] Embodiment 1:

[0054] Compared with other decomposition algorithms, the VSED algorithm has more adjustable parameters, which can control the signal decomposition effect and has good results in rolling bearing fault classification experiments.

[0055] The VSED decomposition is performed on the superposition signal of two sinusoidal signals of different frequencies. The specific steps are as follows:

[0056] C1: First determine the extreme points of the signal to be decomposed and the time corresponding to the extreme points;

[0057] C2: Calculate the value of the long-term trend component at the time corresponding to the extreme point

[0058] C3: Calculate the long-term trend component between the corresponding moments of two extreme points;

[0059] C4: Use the signal to be decomposed to subtract the long-term trend component to obtain the short-term trend component and complete a decomposition.

[0060] Figure 1 The long-term trend component signal diagram of the decomposition of two sinusoidal signals with frequencies of 0.25MHz and 7.5MHz.

[0061] Embodiment 2:

[0062] The SR system can absorb the energy of high-frequency noise in the signal to amplify the weak fault signal in the signal, highlight the characteristics of rolling bearing faults, and ensure the classification accuracy of ELM. The ELM method can obtain the optimal solution with only one calculation, and has the characteristics of fast running speed and good generalization performance.

[0063] like Figure 2 As shown in FIG. 1 , a rolling bearing fault classification method based on variable scale energy decomposition and SR is shown in FIG. 1 . The specific steps are as follows:

[0064] A1: Use the VSED algorithm to decompose the rolling bearing vibration signal and obtain several short-term trend components;

[0065] A2: Calculate the frequency values ​​of several short-term trend components and select the component with a larger fault frequency;

[0066] A3: Use SR to process the components with larger fault frequencies, absorb the noise therein and amplify the fault frequency amplitude;

[0067] A4: The signal obtained in step A3 is randomly divided into data sets, and ELM is used to classify rolling bearing faults to realize the identification of rolling bearing faults.

[0068] Furthermore, the principle of the VSED algorithm in step A1 is as follows:

[0069] Given a signal to be decomposed , define a signal extraction operator , which is used to extract low-frequency information from the original signal to be decomposed and retain it in the long-term trend component. The remaining component is the short-term trend component containing high-frequency information, as shown in the following formula:

[0070]

[0071] in, is the long-term trend component, is the remaining short-term trend component.

[0072] Assumptions For a time series real-valued signal, use Represents a real-valued signal For convenience, we define In the interval Definition and , and in the interval superior Available, so define a signal extraction operator , in the continuous local extreme point interval The Gaussian filter template is generated using the probability density function of the normal distribution , the density function of the normal distribution is as follows:

[0073]

[0074] Baseline signal extraction is performed according to the following formula:

[0075]

[0076] is the Gaussian filter function, according to , Get a Gaussian filter template with a window size of 3 Filtering is performed. ,make ,and .available ,Right now .

[0077] Secondly .and Calculated according to the following formula:

[0078]

[0079] in, , usually taken .and Calculated according to the following formula:

[0080]

[0081] in, , usually taken .

[0082] After a decomposition, a long-term trend component is obtained. Defining the short-term trend component , as shown below:

[0083]

[0084] This decomposition algorithm can extract signals according to different frequencies. The long-term trend component often contains low-frequency information, while the high-frequency information is retained in the short-term trend component. If you want to get multiple components, you can perform multiple decompositions.

[0085] Furthermore, the SR system used in the SR noise reduction process in step A3 is:

[0086]

[0087] Among them, the parameters , is the input signal of the above SR system, is the output signal of the above SR system.

[0088] Furthermore, in step A4, the output signal of SR As input to ELM

[0089] Among them, the principle of ELM is as follows:

[0090] ELM is a typical single hidden layer feedforward neural network. Its network structure consists of input layer, hidden layer and output layer. Assume that the number of neurons in the hidden layer is The input layer has neurons, and the output layer has neurons, assuming The training set input matrix of samples And the output matrix They are:

[0091]

[0092] Assume that the hidden layer neuron activation function is , the network output matrix for:

[0093]

[0094] Among them, ω represents the connection weight between the input layer and the hidden layer, is the neuron threshold of the hidden layer,

[0095] The above formula can be expressed as:

[0096]

[0097] in, The output matrix The transposed matrix of is the hidden layer output matrix, and its specific form is:

[0098]

[0099] When the activation function When infinitely differentiable, ω and It can be randomly determined before network training and remains unchanged during network training. It can be obtained by solving the least squares solution of the following system of equations;

[0100]

[0101] Its solution is

[0102] in is the hidden layer output matrix The generalized inverse matrix of .

[0103] Embodiment 3:

[0104] The experimental data of the present invention adopts the rolling bearing data of Case Western Reserve University in the United States, with a sampling frequency of 12KHz. The rolling bearing state selects a normal state and nine different fault data. The fault data includes inner ring, outer ring and rolling element faults, and three fault data with different fault diameters are selected, a total of nine fault data. Each type of fault data has 100 samples, each sample has 1000 sampling points, 80 samples are used as training sets, and 20 samples are used as test sets. Ten types of faults, a total of 1000 samples. The experimental conditions and sample numbers are shown in Table 1.

[0105] Table 1 Experimental conditions and sample numbers

[0106]

[0107] The specific steps are as follows:

[0108] D1: First, the VSED algorithm is used to decompose the rolling bearing vibration signal four times, and the four short-term trend components obtained are as follows: Figure 3 As shown;

[0109] D2: Perform Fourier transform on these four short-term trend components, calculate their frequencies, and select the component with the larger fault frequency amplitude;

[0110] D3: Perform SR processing on the selected components to absorb the noise energy in the signal and amplify the tiny fault signal in the signal;

[0111] D4: Input the data output by the SR system into the ELM classifier for state identification.

[0112] Among them, label 1 is normal, labels 2, 3, and 4 are respectively the inner ring, rolling element, and outer ring fault states with a fault diameter of 0.1778mm; labels 5, 6, and 7 are respectively the inner ring, rolling element, and outer ring fault states with a fault diameter of 0.3556mm; labels 8, 9, and 10 are respectively the inner ring, rolling element, and outer ring fault states with a fault diameter of 0.5334mm; based on variable scale energy decomposition and SR method processing, the average accuracy of 20 identifications is 99.875%, and the comparison of the results of one classification test set is shown in the figure below. Figure 4 As shown. In this experiment, the number of neurons in the ELM hidden layer is set to 30, and the fault classification accuracy can reach 100%. The classification accuracy of 20 times is shown in Table 2:

[0113] Table 2: Classification accuracy results for 20 consecutive times

[0114]

[0115] As can be seen from the above table, the present invention can effectively enhance the weak fault signal of rolling bearing fault data, ensure the performance of the classifier, and effectively identify the state of the rolling bearing. Compared with the prior art, the present invention has a higher classification accuracy and realizes high-accuracy identification of fault data.

[0116] Three specific examples of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations are made to the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A rolling bearing fault classification method based on variable scale energy decomposition and SR, characterized by: The specific steps are as follows: A1: First, variable scale energy decomposition (VSED) is performed on the actual vibration signal of the rolling bearing to obtain several short-term trend components and one long-term trend component; A 2: Calculate the frequency of each short-term trend component and select the component with the largest fault frequency amplitude; A 3: Use SR to perform noise reduction on the components selected in step A 2; A 4: Data division is performed on the output signal of the SR system, and different fault data are classified using ELM to identify the rolling bearing status.

2. The rolling bearing fault classification method based on variable scale energy decomposition and SR according to claim 1 is characterized in that: The principle of the variable scale energy decomposition algorithm in step A1 is as follows: Given a signal to be decomposed , define a signal extraction operator , which is used to extract low-frequency information from the original signal to be decomposed and retain it in the long-term trend component. The remaining component is the short-term trend component containing high-frequency information, as shown in the following formula: in, is the long-term trend component, is the remaining short-term trend component; Assumptions For a time series real-valued signal, use Represents a real-valued signal For convenience, we define , in the interval Definition and , and in the interval superior Available, so define a signal extraction operator , in the continuous local extreme point interval The Gaussian filter template is generated using the probability density function of the normal distribution , the density function of the normal distribution is as follows: The baseline signal was extracted according to the following formula: is the Gaussian filter function, according to , , get the Gaussian filter template with a window size of 3 Filtering is performed, where ,make ,and ,available ,Right now , Secondly ,and Calculated according to the following formula: in, , usually taken ,and Calculated according to the following formula: in, , usually taken , After a decomposition, a long-term trend component is obtained. , defining the short-term trend component , as shown below: This decomposition algorithm can extract signals according to different frequencies. The long-term trend component often contains low-frequency information, while the high-frequency information is retained in the short-term trend component. If you want to obtain multiple components, you can perform multiple decompositions. The specific steps of the VSED algorithm are as follows: B1: Calculate the extreme point position of the sequence to be decomposed, and calculate the long-term trend component in the sequence to be decomposed. The value of the extreme point position ; B2: Calculate the values ​​between extreme points in the long-term trend component. When the value of the next point is , in the calculation When the value of the previous point is ; B3: Generate a Gaussian filter template using the normal distribution function, set the sum of the weights of the Gaussian filter template to 1, adjust the weights of the first and last values ​​of the template, and filter the values ​​between the extreme points calculated in step B2 to complete the calculation of the long-term trend component; B4: Subtract the long-term trend component obtained in step B3 from the original signal to be decomposed, and the remaining component is the short-term trend component.

3. The rolling bearing fault classification method based on variable scale energy decomposition and SR according to claim 1 is characterized in that: The VSED algorithm in step A1 is combined with SR to construct a method for processing rolling bearing signals; the VSED algorithm is used to extract components containing fault frequencies; and SR is used to amplify weak fault signals.