A method for predicting the remaining life of rolling bearings
Through signal coefficient decomposition and phase space reconstruction, combined with multi-scale fusion arrangement entropy feature extraction and local linear embedding methods, a neural network model for multi-scale feature learning is built, which solves the problem of insufficient expression of noise sensitivity and time-related features in rolling bearing prediction. Achieve higher prediction accuracy and accuracy.
Patent Information
- Application Number
- CN202210608226.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-31
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing method of remaining life prediction of rolling bearings is sensitive to noise components and has insufficient time-related feature expression capabilities, resulting in low prediction accuracy. In particular, the limitations of deep learning learning features at a single scale affect the prediction results.
The signal coefficient decomposition and phase space reconstruction method are used to extract features, and the multi-scale entropy features are extracted through the multi-scale fusion arrangement entropy method. The local linear embedding method is used to remove redundant features to build a neural network model with multi-scale feature learning ability. With attention Force mechanism is used to predict.
The accuracy of the remaining life prediction of rolling bearings is improved, and the life decay feature information can be better captured, the prediction deviation is reduced, and the model fitting ability and generalization ability are enhanced.
Smart Images

Figure CN114993680B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of rolling bearing life prediction, and in particular to a method for predicting the remaining life of a rolling bearing. Background Art
[0002] Rolling bearings are key components included in most mechanical equipment. The health of mechanical equipment is closely related to the health of rolling bearings. Researching the remaining life prediction method of rolling bearings can prevent mechanical equipment failures that lead to downtime or casualties, which is crucial to improving the working efficiency of mechanical equipment.
[0003] In actual engineering applications, in order to better apply bearing vibration signals to subsequent fault diagnosis or remaining life prediction tasks, it is necessary to extract feature information that can reflect its characteristics from the rolling bearing vibration signals. In order to achieve higher fault diagnosis and remaining life prediction accuracy, feature enhancement methods are used to process the rolling bearing feature information. However, the existing feature enhancement methods are sensitive to noise components, or due to the complexity of rolling bearing signals, the ability to express time-related features is not strong, resulting in important feature information not being obvious.
[0004] Remaining life trend prediction usually uses statistical principles, data, and deep learning. The former two have problems with data fitting ability and weak feature learning, which will directly affect the prediction accuracy. Deep learning usually uses learning features at a single scale and has limitations, and there are also problems that affect the prediction accuracy. Summary of the invention
[0005] In order to solve the above problems, the present invention provides a method for predicting the remaining life of a rolling bearing, which can reflect more trends of the remaining life of the bearing changing over time and has good remaining life prediction accuracy.
[0006] The present invention provides a method for predicting the remaining life of a rolling bearing, comprising the following steps:
[0007] S1, extract the original vibration signal and perform noise reduction processing;
[0008] S2, decomposing and reconstructing the denoised signal into a short-time sequence matrix, and extracting features by a multi-scale fusion permutation entropy method;
[0009] S3. Input the extracted feature data into the neural network to obtain the predicted remaining life trend of the rolling bearing.
[0010] Further, step S2 includes:
[0011] S21, decomposing the low resonance component X in the noise-reduced signal;
[0012] S22, reconstructing the low resonance component X into several short-time sequences, thereby obtaining a multi-dimensional short-time sequence matrix: Y = {Y1, Y2, Y3, ...Y N};
[0013] S23, obtaining a grayscale image by performing time delay segmentation and phase space sparse reconstruction on the short-time sequence, and calculating the probability distribution of all pixels in the grayscale image;
[0014] S24. Obtain a multi-scale permutation entropy value of each short-time sequence according to the probability distribution of the pixel points, and then obtain a multi-scale permutation entropy value matrix, perform dimensionality reduction and fusion processing on the multi-scale permutation entropy value matrix, and obtain a multi-scale fusion permutation entropy feature.
[0015] Furthermore, the decomposition method of step S21 adopts a resonance sparse decomposition method.
[0016] Furthermore, the reconstruction method of step S22 adopts a sliding window slicing method.
[0017] Further, step S23 includes:
[0018] S231, each short-time sequence Y i Reconstructed into multidimensional matrix A i ;
[0019] S232, the multidimensional matrix A i Normalization is performed and the values at the corresponding positions of the matrix are converted into corresponding pixel points to obtain the grayscale image A′. i ;
[0020] S233, using a histogram equalization method to enhance the grayscale image A′ i Feature information in ;
[0021] S234, calculating the grayscale image A′ i The pixel probability distribution H i .
[0022] Further, step S24 includes:
[0023] S241, calculating the multi-scale permutation entropy of the pixel probability distribution, and then obtaining a multi-scale permutation entropy value matrix;
[0024] S242. Use a local linear embedding method to reduce the dimension of the multi-scale permutation entropy value matrix to obtain a multi-scale fusion permutation entropy feature.
[0025] Further, step S242 includes:
[0026] S2421, finding the entropy value e in the multi-scale permutation entropy value matrix according to the Euclidean distance iThe k nearest neighbors of i Linear relationship with k nearest neighbors;
[0027] S2422, calculate the covariance matrix Z in the k nearest neighbor sample space i , and then get the corresponding weight coefficient vector W i ;
[0028] S2423, according to the weight coefficient vector W i With the weight coefficient matrix W, we can calculate the condition matrix M according to the following formula:
[0029] M=(1-W)(1-W) T ;
[0030] S2424. Calculate the first d+1 eigenvalues of the conditional matrix M, and calculate the eigenvectors corresponding to the d+1 eigenvalues; a multi-scale fusion permutation entropy matrix of the matrix composed of the second eigenvector to the d+1th eigenvector.
[0031] Furthermore, the neural network includes a feature extraction module and an attention mechanism module, and the feature information output by the feature extraction module is fused by the attention mechanism module to obtain a remaining life prediction value.
[0032] Furthermore, the feature extraction module includes several parallel convolution modules.
[0033] Furthermore, the feature extraction module includes a first convolution module, a second convolution module, a third convolution module and a fourth convolution module, the convolution kernel size of the first convolution module is (1, 1), the convolution kernel of the second convolution module is (1, 2), the convolution kernel of the third convolution module is (1, 3), and the convolution kernel of the fourth convolution module is (1, 4).
[0034] As described above, compared with the prior art, the present invention has the following effects:
[0035] 1. The present invention adopts the method of signal coefficient decomposition and space reconstruction to decompose and reconstruct the signal, enhance the characteristic information in the signal, and adopts the feature extraction method of multi-scale fusion permutation entropy. After extracting the multi-scale entropy feature matrix, the local linear embedding method is used to remove redundant feature information, so that the extracted features can carefully reflect the decay trend of the remaining life of the rolling bearing.
[0036] 2. The present invention builds a neural network model with multi-scale feature learning capabilities, which can extract the feature information of MFPE features at different time scales, integrate multi-scale features, improve the model's fitting ability, and reduce prediction deviations. Comparative experiments have shown that the prediction error of the neural network model of the present invention is within 5 minutes, which can better capture the life decay feature information and has good remaining life prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 It is a schematic diagram of the overall process of a specific embodiment of the present invention;
[0038] Figure 2 A schematic diagram of a neural network structure of a specific embodiment of the present invention;
[0039] Figure 3 are the high resonance component and the low resonance component of the signal obtained after the signal sparse decomposition processing in the specific embodiment of the present invention, Figure 3 a is the high resonance component, Figure 3 b is the low resonance component;
[0040] Figure 4 The grayscale image obtained by using the phase space reconstruction method for the low resonance component in the specific embodiment of the present invention is: Figure 4 a is the original grayscale image, Figure 4 b is the grayscale image processed by the histogram equalization method;
[0041] Figure 5 is a multi-scale permutation entropy matrix of a specific embodiment of the present invention;
[0042] Figure 6 is the MFPE characteristic curve of a specific embodiment of the present invention;
[0043] Figure 7 This is a graph showing the remaining life prediction result of a rolling bearing according to a specific embodiment of the present invention. Figure 7 a is the remaining life prediction result obtained by using the MACNN model of the present invention, Figure 7 b is the remaining life prediction result obtained by using the CNN model; Figure 7 c is the remaining life prediction result obtained by using the MCNN model, Figure 7 d is the remaining life prediction result obtained using the Attention-CNN model. DETAILED DESCRIPTION
[0044] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.
[0045] It should be noted that the illustrations provided in the following embodiments are only used to illustrate the basic concept of the present invention in a schematic manner, and thus the illustrations only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0046] like Figure 1 As shown, in a specific embodiment, a method for predicting the remaining life of a rolling bearing is provided, comprising the following steps:
[0047] S1, extract the original vibration signal and perform noise reduction processing;
[0048] The noise reduction process adopts the EEMD-WSST method, which specifically includes:
[0049] Firstly, the original vibration signal is decomposed into IMF components by the EEMD method. The IMF components with more noise components are screened out using the Pearson correlation analysis. Then, the IMF components are screened out using the WSST method for denoising. While ensuring a good signal denoising effect, the characteristic information in the signal is preserved, thereby improving the effectiveness of subsequent feature extraction methods.
[0050] S2, decomposing and reconstructing the denoised signal into a short-time sequence matrix, and extracting features by a multi-scale fusion permutation entropy method;
[0051] Feature enhancement methods can improve the expressiveness of features, and the sensitivity of features to the decline trend of the remaining life of rolling bearings directly affects the accuracy of the prediction model.
[0052] S21, using a resonance coefficient decomposition method to decompose a low resonance component X in the noise-reduced signal;
[0053] The specific calculation steps of the resonance sparse decomposition algorithm of this embodiment are as follows:
[0054] Input signal X=X′+X, set the resonance sparse decomposition parameter quality factors Q1, Q2, redundancy factors r1, r2 and decomposition levels J1, J2 according to the signal characteristics, and construct the wavelet basis function library S1, S2.
[0055] According to the signal-to-noise ratio index of the signal, select appropriate weighting coefficients λ1 and λ2. In order to effectively separate the different components in the signal, set the optimization target as shown in the following formula:
[0056]
[0057] Where: ω1, ω2 are the matching coefficients of wavelet bases S1 and S2.
[0058] By solving the optimization problem of the above formula, the optimal matching coefficient is obtained By the best matching coefficient The high resonance component X′ and the low resonance component X are obtained by combining with the wavelet basis function library.
[0059] S22, reconstruct the low resonance component X into several short-time sequences by using a sliding window slicing method, and then obtain a multi-dimensional short-time sequence matrix: Y = {Y1, Y2, Y3, ...Y z}, where Y i represents the i-th short-time series, specifically:
[0060] Y i =[x((i-1)s+1),x((i-1)s+2),...,x((i-1)s+Nω)] T
[0061]
[0062] Wherein, z is the number of short-time sequences, Nω is the length of the short-time sequence, s is the delay parameter, and s∈(1,Nω).
[0063] The signal is sampled with overlap through a sliding window of fixed length, and the overall signal is converted into a short-time series with mutual connections. When the multi-scale entropy feature extraction is performed later, the local feature information in the signal can be better reflected. The signal is converted from an overall long-time series into multiple short-time series. According to the different delay parameters selected, different overlapping phenomena will appear in multiple short-time series, which can further enrich the fault information contained in the enhanced signal. When making the remaining life prediction, the rolling bearing full life cycle signal is used. The amount of data contained in the signal will affect the calculation time and prediction time of the prediction model. After the sliding window slicing processing, the amount of calculation for each model iteration is reduced, and the difficulty and complexity of the model calculation are reduced.
[0064] S23, obtaining a grayscale image by performing time delay segmentation and phase space sparse reconstruction on the short-time sequence, and calculating the probability distribution of all pixels in the grayscale image, specifically including:
[0065] S231, use the Hankel matrix to transform each short-time series Y iReconstructed into m×n multidimensional matrix A i , as shown below:
[0066]
[0067] Where m, n satisfy m+m+1=Nω;
[0068] S232, the multidimensional matrix A i Normalization is performed and the values at the corresponding positions of the matrix are converted into corresponding pixel points to obtain the grayscale image A′. i ;
[0069] S233, using a histogram equalization method to enhance the grayscale image A′ i Feature information in ;
[0070] S234, calculating the grayscale image A′ i The pixel probability distribution H i :
[0071] H i =[h 1,i ,h 2,i ,...,h 256,i ] T ;
[0072] In this embodiment, by reconstructing the signal in phase space, each short-time signal sequence is constructed as a two-dimensional matrix and normalized, and the data in the matrix is converted into a grayscale image as pixel values. By changing the distribution of the pixel histogram in the image, the expression of the feature information in the image is enhanced, and the histogram algorithm does not change the feature trend in the image. In essence, it is equivalent to increasing the pixel value of the data point where the feature information is located, making the feature information in the image more prominent, further enhancing the feature information in the signal.
[0073] S24. Obtain a multi-scale permutation entropy value of each short-time sequence according to the probability distribution, and then obtain a multi-scale permutation entropy value matrix, perform dimensionality reduction and fusion processing on the multi-scale permutation entropy value matrix, and obtain a multi-scale fusion permutation entropy feature.
[0074] S241, calculating the multi-scale permutation entropy of the pixel probability distribution, and then obtaining a multi-scale permutation entropy value matrix;
[0075] Permutation entropy has good characterization capabilities for different faults occurring in rolling bearings, but permutation entropy can only evaluate the characteristic information of rolling bearing vibration signals on a single time scale, which may cause key characteristic information to be unclear. At present, the vibration signal of rolling bearings is a complex signal, and evaluating it only from a single time scale can no longer fully reflect the complete characteristic information. Therefore, this embodiment uses multi-scale permutation entropy to improve the shortcomings of permutation entropy. Multi-scale permutation entropy can reflect deeper features in the signal by setting different scale factors, so as to solve the shortcomings of permutation entropy. The specific method for calculating multi-scale permutation entropy includes:
[0076] With grayscale probability distribution H = {H1, H2, ..., H z} as input, and coarse-grained it by the following formula to obtain the coarse-grained sequence
[0077]
[0078] Where τ is the scale factor;
[0079] The coarse-grained sequence Phase space reconstruction obtains the multi-scale sequence Y i (τ) , as shown below:
[0080]
[0081] Where Y i (τ) is a multi-scale sequence, m is the embedding dimension, and s is the time delay sparsity.
[0082] The multi-scale time series Y i (τ) Arrange in ascending order, record the index θ of each short-time series after ascending order j ={j1,j2,...,j m}, each short-time sequence has m! kinds of arrangement, and the number of times each arrangement occurs is counted N l , and calculate the frequency of each arrangement as shown below:
[0083]
[0084] Calculate the permutation entropy of multi-scale time series to get the multi-scale permutation entropy E i :
[0085]
[0086] The multi-scale permutation entropy matrix MPEM can be obtained:
[0087]
[0088] S242. Use a local linear embedding method to reduce the dimension of the multi-scale permutation entropy value matrix to obtain a multi-scale fusion permutation entropy feature.
[0089] Multiscale permutation entropy reconstructs the signal through coarse-graining and phase space reconstruction operations, and can obtain the characteristic information of the signal at different time scales, which improves the problem of incomplete characteristic information in a single dimension and can improve the remaining life prediction accuracy. Since the short-time series matrix is constructed using the sliding window slicing processing method, the signal partially overlaps. Although this operation can enrich the characteristic information in the signal, it will also cause redundancy of features in the signal that are insensitive to the remaining life decay trend of the rolling bearing, thereby causing feature redundancy in the high-dimensional multiscale permutation entropy feature matrix.
[0090] The characteristic information contained in the rolling bearing vibration signal is not all sensitive to the trend of the remaining life of the rolling bearing declining over time. When using multi-scale permutation entropy as an evaluation feature, as the scale factor increases, the dimension of the multi-scale permutation entropy value also increases. The multi-scale permutation entropy value will inevitably contain insensitive characteristic information, which affects the accuracy of the subsequent prediction of the remaining life of the rolling bearing. Therefore, it is necessary to reduce the dimensionality of the extracted multi-scale permutation entropy features to remove the redundant and insensitive characteristic information.
[0091] S2421, finding the entropy value e in the multi-scale permutation entropy value matrix according to the Euclidean distance i The k nearest neighbors of i The linear relationship between the k nearest neighbors, the loss function is:
[0092]
[0093] In the formula, K(i) represents the entropy value e i The k nearest neighbor sample set of ω; ij is the linear weight coefficient, which is generally normalized to meet the following conditions. i The entropy value of the k nearest neighbor sample set will make its weight coefficient 0, and expand the weight coefficient to the dimension of the entire data set.
[0094]
[0095] S2422, calculate the covariance matrix Z in the k nearest neighbor sample space i , and then get the corresponding weight coefficient vector W i ;
[0096] Z i =(x i -x j )(xi -x j ) T ;
[0097]
[0098] In the formula, 1 k is a k-dimensional vector with value 1.
[0099] S2423, according to the weight coefficient vector W i With the weight coefficient matrix W, we can calculate the condition matrix M according to the following formula:
[0100] M=(1-W)(1-W) T ;
[0101] In the formula, I is the constraint condition to ensure that the entropy value retains the original feature information as much as possible after dimensionality reduction. y i is the fusion entropy value obtained after dimensionality reduction.
[0102] S2424, calculate the first d+1 eigenvalues of the conditional matrix M, and calculate the eigenvectors {y1, y2, ..., y d+1}, the matrix composed of the second eigenvector to the d+1th eigenvector is the multi-scale fusion permutation entropy matrix MFPE = {y2, y3, ..., y d+1}.
[0103] The high-dimensional multi-scale permutation entropy value matrix is reduced to a one-dimensional vector through the above method to obtain a multi-scale fusion permutation entropy feature vector reflecting the remaining life decay trend of the bearing. The multi-scale fusion permutation entropy feature vector is input into the subsequent judgment model to obtain the remaining life trend of the ball bearing.
[0104] S3. Input the extracted feature data into the neural network to obtain the predicted remaining life trend of the rolling bearing.
[0105] When a convolutional neural network extracts features from input features at a single scale, it often cannot reflect the feature information well. Therefore, a multi-scale convolutional neural network is proposed as a basic model for predicting the remaining life of rolling bearings. Figure 2 As shown, it includes a feature extraction module and an attention mechanism module, the feature extraction module includes a first convolution module, a second convolution module, a third convolution module and a fourth convolution module, the convolution kernel size of the first convolution module is (1, 1), the convolution kernel of the second convolution module is (1, 2), the convolution kernel of the third convolution module is (1, 3), and the convolution kernel of the fourth convolution module is (1, 4). The feature information output by the feature extraction module is fused by the attention mechanism module to obtain the remaining life prediction value.
[0106] This embodiment builds a MACNN rolling bearing remaining life prediction model based on a multi-scale feature extraction module and an Attention mechanism. The overall model first uses a convolution kernel of size (1, 1) to extract shallow features of the input data, and then uses four convolution modules to extract deep features at different scales. Because the overall model has many parameters, in order to prevent the model from overfitting, the shallow features are spliced with the extracted deep features using a residual connection method. The spliced multi-scale features are input into the Attention fusion layer to obtain the fused Attention feature vector, which is then input into the fully connected layer to obtain the final prediction result.
[0107] In order to further verify the above-mentioned embodiment of the present invention, in a specific embodiment, the XJTU-SY bearing data set is used for experimental verification. The XJTU-SY bearing data information table is shown in Table 1. In the experimental verification, Bearing1_1 data is used as the verification set, and Bearing1_2 is used as the training set.
[0108] Table 1: XJTU-SY bearing data information table
[0109]
[0110] The data is denoised using EEMD-WSST, and the denoised data is subjected to resonance sparse decomposition to obtain Figure 3 a and Figure 3 b. As can be seen from the figure, both the low resonance component and the high resonance component can well preserve the overall trend information of the original signal. Throughout the life cycle, the characteristic information in the low resonance component is more obvious, and the characteristic information of the high resonance component is mostly submerged. This proves that the low resonance component contains more characteristic information. As can be seen from the figure, although the overall amplitude of the low resonance component obtained after resonance sparse decomposition decreases, the characteristic information in the signal is more prominent, and the ratio of the frequency component in the frequency range of 200Hz-1000Hz to the highest frequency component becomes smaller, which can better extract the characteristic information contained in these frequency bands.
[0111] The low resonance components of the resonance sparse decomposition are selected and constructed into a short-time series matrix through the sliding window slicing processing method, and the phase space reconstruction method is used to enhance its internal features, such as Figure 4 As shown. Figure 4 It can be seen that compared with Figure 4 The original grayscale image shown in a and the grayscale image processed by the histogram equalization method ( Figure 4 The characteristic information in b) is more obvious, and the stripes therein represent the characteristic information inside thereof, which proves the effectiveness of the above-mentioned feature enhancement method in the above-mentioned embodiment.
[0112] Then the pixel probability distribution of the image after phase space reconstruction is calculated, and then the multi-scale permutation entropy value of the pixel probability distribution is calculated to obtain the multi-scale permutation entropy matrix, such as Figure 5 As the remaining life of the bearing declines, that is, the bearing gradually degrades, the multi-scale arrangement entropy value shows an obvious upward trend, and there is an obvious step-like increase, which corresponds to the step-like increase in the amplitude of the bearing vibration signal, which can well reflect the serious degradation of the bearing.
[0113] The LLE dimensionality reduction algorithm is used to fuse the high-dimensional entropy matrix to obtain the MFPE feature, such as Figure 6 As shown in the figure, the MFPE feature can reduce feature redundancy while retaining the main remaining life trend information in the multi-scale permutation entropy matrix. When the bearing is severely degraded, the MFPE feature also undergoes a step improvement.
[0114] A MACNN bearing life model is built to predict the remaining life trend of the bearing. To meet the model input and training requirements, in this embodiment, the MFPE feature matrix size is reconstructed to (992, 1, 12), and the MFPE features are input into the built MACNN life prediction model to obtain the prediction results as shown below: Figure 7 As shown in a, for the convenience of comparison, the data set is input into the existing CNN model, MCNN model, and Attention-CNN model, and the output prediction results are as follows Figure 7 b. Figure 7 c. Figure 7 d, the prediction performance indicators of each model are shown in Table 2:
[0115] Table 2: Prediction performance indicators of different models
[0116]
[0117] from Figure 7 It can be seen from the prediction results that the overall life prediction effect of the MACNN model is better, the overall trend is the same as the real life curve, and the life prediction value deviates less from the real value. Compared with the MACNN life prediction model, the MCNN model has a deviation between the real trend and the predicted trend at the end of the bearing life, and the overall model deviates greatly from the real value; the CNN model and the Attention-CNN model fail to reflect the real bearing life trend at the end of the bearing life, and the overall life prediction value deviates greatly from the real life value. According to the comparison of life prediction results between different models, it can be seen that after adding the multi-scale feature extraction module, the overall performance of the CNN model has been greatly improved. The addition of the Attention mechanism can make the life prediction results closer to the real value and improve the accuracy of life prediction.
[0118] The three indicators of MAE, RMSE and model score are used to quantitatively analyze the model prediction results, as shown in Table 5-3. As can be seen from Table 5-3, the total score of the MACNN model is improved by 13.17%, 51.01% and 25.36% compared with the MCNN model, CNN model and Attention-CNN model, respectively, indicating that the MACNN model has better prediction accuracy and comprehensive performance in the prediction of remaining life; the MAE evaluation index is reduced by 9.91%, 37.41% and 32.5% compared with the MCNN model, CNN model and Attention-CNN model, respectively, indicating that the MACNN model has better fitting ability; the RMSE evaluation index is reduced by 15.03%, 41.98% and 38.02% compared with the MCNN model, CNN model and Attention-CNN model, respectively, indicating that the MACNN model has better generalization ability.
[0119] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.
Claims
1. A method for predicting the remaining life of a rolling bearing, characterized in that: The steps include: S1, extract the original vibration signal and perform noise reduction processing; S2, decomposing and reconstructing the denoised signal into a short-time sequence matrix, and extracting features by a multi-scale fusion permutation entropy method, specifically including: S21, using a resonance sparse decomposition method to decompose a low resonance component X in the denoised signal; S22, reconstructing the low resonance component X into several short-time sequences, thereby obtaining a multi-dimensional short-time sequence matrix: Y = {Y1, Y2, Y3, ...Y N }; S23, obtaining a grayscale image by performing time delay segmentation and phase space sparse reconstruction on the short-time sequence, and calculating the probability distribution of all pixels in the grayscale image, specifically including: S231, each short-time sequence Y i Reconstruct into multidimensional matrices A i ; S232, the multidimensional matrix A i Normalization is performed and the values at the corresponding positions of the matrix are converted into corresponding pixel points to obtain a grayscale image. ; S233, enhancing the grayscale image using a histogram equalization method Feature information in ; S234, calculating the grayscale image The pixel probability distribution H i ; S24, calculating the pixel probability distribution to obtain the multi-scale permutation entropy value of each short-time sequence, and then obtaining a multi-scale permutation entropy value matrix, performing dimensionality reduction fusion processing on the multi-scale permutation entropy value matrix, and obtaining a multi-scale fusion permutation entropy feature; S3. Input the extracted feature data into the neural network to obtain the predicted remaining life trend of the rolling bearing.
2. A method for predicting the remaining life of a rolling bearing according to claim 1, characterized in that: The reconstruction method of step S22 adopts a sliding window slicing method.
3. A method for predicting the remaining life of a rolling bearing according to claim 1, characterized in that: The dimensionality reduction method of the multi-scale permutation entropy value matrix is a local linear embedding method.
4. A method for predicting the remaining life of a rolling bearing according to claim 3, characterized in that: The method of using the local linear embedding method to perform dimensionality reduction and fusion processing on the multi-scale arrangement entropy value matrix includes: Find the entropy value in the multi-scale permutation entropy value matrix according to the Euclidean distance e i The k nearest neighbors of e i Linear relationship with k nearest neighbors; Calculate the covariance matrix Z in the k nearest neighbor sample space i , and then get the corresponding weight coefficient vector W i ; According to the weight coefficient vector Wi, the weight coefficient matrix W is obtained, and the condition matrix M is calculated according to the formula: M=(I-W)(I-W) T ; Calculate the first d+1 eigenvalues of the conditional matrix M, and calculate the eigenvectors corresponding to the d+1 eigenvalues; a multi-scale fusion permutation entropy matrix of the matrix composed of the second eigenvector to the d+1th eigenvector.
5. A method for predicting the remaining life of a rolling bearing according to claim 1, characterized in that: The neural network includes a feature extraction module and an attention mechanism module. The feature information output by the feature extraction module is fused by the attention mechanism module to obtain a remaining life prediction value.
6. A method for predicting the remaining life of a rolling bearing according to claim 5, characterized in that: The feature extraction module includes several parallel convolution modules.
7. A method for predicting the remaining life of a rolling bearing according to claim 6, characterized in that: The feature extraction module includes a first convolution module, a second convolution module, a third convolution module and a fourth convolution module. The convolution kernel size of the first convolution module is (1, 1), the convolution kernel size of the second convolution module is (1, 2), the convolution kernel size of the third convolution module is (1, 3), and the convolution kernel size of the fourth convolution module is (1, 4).
