Bearing life prediction method based on novel health indexes

Abstract

The invention discloses a bearing life prediction method based on a novel health index, and particularly relates to the technical field of bearing residual life prediction, and the method comprises the following steps: S1, collecting a vibration signal of a rolling bearing through a vibration sensor, and carrying out the preprocessing of the vibration signal; s2, extracting multi-domain features from the preprocessed vibration signals to construct health indexes of the bearing; s3, establishing a health index degradation model, setting a failure threshold value, and then determining prediction of the residual life of the bearing according to probability distribution of time when the equipment degradation amount reaches the failure threshold value for the first time; according to the method, fluctuation caused by noise existence is reduced through completely adaptive noise set empirical mode decomposition and kernel principal component analysis; compared with an existing residual service life prediction method, the method has the advantages that the multi-domain features are extracted, so that the constructed bearing health index contains rich degradation information, and the accuracy of a prediction result is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing remaining life prediction, and particularly relates to a method for predicting bearing life based on a new health index. Background Art

[0002] With the progress of science and technology, mechanical equipment has become more complex and precise. Prediction and health management have become effective solutions to improve the reliability, safety, and maintenance ability of mechanical equipment. Therefore, the research on motor bearing fault prediction is of great significance.

[0003] Previous research on bearing fault prediction mainly relied on a single health index to predict the remaining service life. Currently, many aspects have systematically summarized the research on using a single health index to predict the faults of rolling bearings in recent years. On the one hand, the fluctuations caused by the presence of noise are ignored. On the other hand, the single health index constructed for the bearing lacks rich degradation information.

[0004] For the above reasons, it is necessary to propose a method for predicting bearing faults by extracting multi-domain features and proposing a new health index. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for predicting bearing life based on a new health index, aiming to solve the problems in the prior art that when predicting the remaining service life of rolling bearings relying on a single health index, the fluctuations caused by the presence of noise are ignored and the single health index constructed for the bearing lacks rich degradation information.

[0006] To achieve the above purpose, a method for predicting bearing life based on a new health index is provided. The prediction method includes the following steps:

[0007] S1. Use a vibration sensor to collect the vibration signal of the rolling bearing and preprocess the vibration signal;

[0008] S2. Extract multi-domain features from the preprocessed vibration signal to construct a health index for the bearing;

[0009] S3. Establish a health index degradation model based on the nonlinear Wiener process and combine the failure threshold to achieve high-precision prediction of the bearing remaining life.

[0010] Further, in step 1, the preprocessing of the vibration signal is to process the vibration signal by the method of continuous wavelet transform, and the expression is:

[0011]

[0012] In the formula: x(t) represents a one-dimensional vibration signal, denotes the complex conjugate operation, a is the scaling factor, b is the translation factor, ψ a,b is the mother wavelet, t is the time, and X(a, b) is the time-frequency representation TFR used.

[0013] Furthermore, in step S2, the process of extracting multi-domain features from the preprocessed vibration signal to construct the health index of the bearing includes:

[0014] S2-1: Obtain the non-linear health index HI of the bearing using complete ensemble empirical mode decomposition with adaptive noise and kernel principal component analysis;

[0015] S2-2: Construct multi-health indices HIs in the dual-channel transformer network-convolutional block attention module DCTN-CBAM architecture according to the multi-domain features of the bearing;

[0016] S2-3: Use the constructed multi-health indices HIs of the bearing for the first prediction time FPT detection.

[0017] Furthermore, the complete ensemble empirical mode decomposition process in step S2-1 includes:

[0018] (1) Construct an adaptive noise sequence and add it to the original signal to generate multiple groups of random trial signals;

[0019] (2) Perform complete ensemble empirical mode decomposition with adaptive noise CEEMD on each group of random trial signals to obtain a group of intrinsic mode function IMF functions;

[0020] (3) Combine and weight each group of intrinsic mode function IMF functions to generate a group of total intrinsic mode function IMF functions;

[0021] (4) Perform complete ensemble empirical mode decomposition CEEMD decomposition on the total intrinsic mode function IMF functions to obtain a group of new intrinsic mode function IMF functions;

[0022] (5) Check whether the newly generated intrinsic mode function IMF functions meet the convergence condition. If they do not meet the convergence condition, go back to step (1) to adjust the adaptive noise;

[0023] (6) If the convergence condition is met, then repeat steps (2) to (5) for iteration until the number of obtained IMF functions no longer increases.

[0024] Furthermore, in step 2-1, the kernel principal component analysis process:

[0025] (1) Calculate the kernel matrix K:

[0026] K ij = k(x i , x j);

[0027] Where x i and x j are both sample points in the dataset;

[0028] (2) Center the kernel matrix:

[0029] K′ = K - 1 n K - K1 n + 1 n K1 n

[0030] Where 1 n is a matrix of all 1s, used to eliminate the influence of the mean;

[0031] (3) Calculate the eigenvalues and eigenvectors of the kernel matrix:

[0032] K′v i = λ i v i ;

[0033] Where λ i is the eigenvalue of the kernel matrix, and v i is the corresponding eigenvector;

[0034] (4) Select the eigenvectors corresponding to the top K largest eigenvalues as the new reduced-dimensional space.

[0035] Furthermore, the dual-channel transformer network-convolutional block attention module DCTN-CBAM architecture described in step S2-2 consists of a dual-channel transformer network and a convolutional block attention module. The dual-channel transformer network: one channel is used to process time-domain and frequency-domain features, and the other channel is used to extract degradation information from the reduced time-frequency representation TFR. The convolutional block attention module includes a channel attention module and a spatial attention module.

[0036] Furthermore, the channel attention module is expressed as:

[0037]

[0038] Where W0 and W1 are weight matrices, are the features after average pooling and max pooling, S is the input feature, S′ is the feature map after the channel attention module, and M c refers to performing the channel attention module operation on the input feature.

[0039] Furthermore, the spatial attention module is expressed as:

[0040]

[0041] Where M sIt refers to performing a spatial attention module operation on the input features, f n×n (·) represents a convolution operation with a kernel size of n, represents the stacked feature maps.

[0042] Furthermore, the process of the first prediction time FPT detection described in step S2-3 includes: using the bearing multi-health indicators HIs to determine the first prediction time FPT, and using 3σ to construct the monitoring interval of the health state. The first prediction time FPT is defined as follows:

[0043] |HI t+i -m| > 3σ s (i = 0, 1, 2)

[0044] In the formula, HI t+i is the (t + i)-th value of the health indicator HI, and m and σ s are the mean and standard deviation of the health state monitoring interval respectively.

[0045] Furthermore, in step S3, the process of establishing a health indicator degradation model, setting a failure threshold, and then determining the prediction of the remaining life of the bearing according to the probability distribution of the time when the equipment degradation amount first reaches the failure threshold includes:

[0046] (1) For the non-linear characteristics of the rolling bearing health indicator HI, a non-linear Wiener process is used to characterize the degradation characteristics of the health indicator HI; let {X(t), t > 0} be the degradation process of the health indicator HI, and the non-linear Wiener process is defined as follows:

[0047] X(t) = X(0) + at b + σ B B(t)

[0048] In the formula, a is the drift parameter, σ B is the diffusion parameter, b is the unknown parameter in the non-linear function of the drift coefficient, and B(t) represents the standard Brownian motion process reflecting the random dynamics of the degradation process and the inherent time-varying uncertainty used to describe the performance degradation process;

[0049] (2) A failure threshold is predefined. The life of the rolling bearing ends when its health indicator HI exceeds the failure threshold, and the probability prediction of the remaining service life is carried out. The process is as follows:

[0050] First, define the failure threshold C, which is defined as:

[0051] T = inf{t > 0: X(t) ≥ C|t > 0}

[0052] In the formula, T is the remaining service life, defined as the time when the health indicator first exceeds the failure threshold C;

[0053] Then calculate the probability distribution of the time when the equipment degradation amount first reaches the failure threshold (i.e., the remaining life), and its expression is:

[0054]

[0055] In the formula, σ B is the diffusion parameter, b is the unknown parameter in the drift coefficient non-linear function, C is the failure threshold, σ a is the standard deviation of the drift parameter a, μ a is the mean value of the drift parameter a;

[0056] The remaining service life RUL of the rolling bearing at the k-th observation time is represented by L k as:

[0057] L k =inf{l k >0:X(l k +t k )≥C|X(t k )<C};

[0058] In the formula, t k is the time point of the k-th observation, and L k is the remaining service life at the time t k of the k-th observation.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] 1. The present invention reduces the fluctuations caused by the presence of noise through complete adaptive noise ensemble empirical mode decomposition and kernel principal component analysis.

[0061] 2. Compared with the current remaining service life prediction methods, multi-domain features are extracted, so that the constructed bearing health index contains rich degradation information and improves the accuracy of the prediction results.

[0062] On the basis of the implementation manners provided in the above aspects, the present application can be further combined to provide more implementation manners. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. 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.

[0064] Figure 1 is the flow chart of the bearing remaining life prediction of the present invention;

[0065] Figure 2 is the flow chart for constructing the health indicators of the present invention;

[0066] Figure 3 is the processing process diagram of the convolutional attention module;

[0067] Figure 4 is the comparison diagram of the prediction results of 5 rolling bearings of the same model under different working conditions. Specific implementation manners

[0068] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. If not specifically specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0069] The embodiment of the present invention provides a method for predicting the bearing life based on a new health indicator. Refer to Figure 1 , Figure 1 is the flow chart of a method for predicting the bearing life based on a new health indicator provided by the embodiment of the present invention. First, the vibration signal during the operation of the rolling bearing is collected by a vibration sensor, and the collected vibration signal is processed by the method of continuous wavelet transform; multi-domain features are extracted from the processed vibration signal to construct the health indicator of the bearing. First: the non-linear health indicator HI of the bearing is obtained by using complete adaptive noise ensemble empirical mode decomposition and kernel principal component analysis; subsequently, HIs are constructed in the DCTN-CBAM architecture according to the multi-domain features of the bearing; then, the constructed rolling bearing HIs are used for FPT detection; finally, degradation modeling and remaining useful life probability prediction are performed to obtain the final prediction result. The specific implementation steps are as follows:

[0070] Step 1: Use a vibration sensor to collect the vibration signal of the rolling bearing and preprocess the vibration signal.

[0071] According to the vibration sensor to collect the vibration signal during its operation, the collected vibration signal is processed by the method of continuous wavelet transform, and its expression is:

[0072]

[0073] In the formula: x(t) represents a one-dimensional vibration signal, represents the complex conjugate operation, a is the scaling factor, b is the translation factor, ψ a,bLet \(\psi(t)\) be the mother wavelet, \(t\) be the time, and \(X(a,b)\) be the TFR used.

[0074] Step 2: Extract multi-domain features from the preprocessed vibration signal to construct the health index of the bearing.

[0075] Since a single health index does not contain rich degradation information, a multi-domain feature extraction method is adopted to calculate the time-domain, frequency-domain, and time-frequency-domain features of the vibration signal, forming a multi-dimensional feature vector. A new type of health index is constructed to comprehensively characterize the state of the bearing.

[0076] The process of constructing the health index is as follows: Figure 2 As shown in the specific process of constructing the health index, it is elaborated as follows: First, the preprocessed vibration signal is adaptively decomposed by complete ensemble empirical mode decomposition with adaptive noise to suppress noise interference. Combined with kernel principal component analysis, high-dimensional mapping and dimensionality reduction are performed on the non-linear degradation features to construct the non-linear health index HI. Subsequently, using the DCTN-CBAM architecture, the time-domain statistical features and frequency-domain spectral features of the vibration signal are extracted through parallel convolutional channels respectively, and the composite health index HIs is constructed using the channel attention module and the spatial attention module. Finally, the critical point where the health index exceeds the preset failure threshold is captured by FPT detection, and the 3σ is used to construct the monitoring interval of the health state.

[0077] Specifically, the process of complete ensemble empirical mode decomposition with adaptive noise includes:

[0078] (1) Construct an adaptive noise sequence and add it to the original signal to generate multiple groups of random test signals;

[0079] (2) Perform complete ensemble empirical mode decomposition with adaptive noise (CEEMD) on each group of random test signals to obtain a group of intrinsic mode function (IMF) functions.

[0080] (3) Combine and weight each group of IMF functions to generate a group of total IMF functions.

[0081] (4) Perform complete ensemble empirical mode decomposition with adaptive noise (CEEMD) on the total IMF functions to obtain a new group of IMF functions.

[0082] (5) Check whether the newly generated IMF functions meet the convergence condition. If they do not meet the convergence condition, return to step (1) to adjust the adaptive noise.

[0083] (6) If the convergence condition is met, repeat steps (2) to (5) for iteration until the number of IMF functions obtained no longer increases.

[0084] Specifically, the process of kernel principal component analysis:

[0085] (1) Calculate the kernel matrix K:

[0086] K ij = k(x i , x j )

[0087] where x i and x j are both sample points in the dataset.

[0088] (2) Center the kernel matrix:

[0089] K′ = K - 1 n K - K1 n + 1 n K1 n

[0090] where 1 n is a matrix of all 1s, used to eliminate the influence of the mean.

[0091] (3) Calculate the eigenvalues and eigenvectors of the kernel matrix:

[0092] K′v i = λ i v i .

[0093] where λ i is the eigenvalue of the kernel matrix, and v i is the corresponding eigenvector.

[0094] (4) Select the eigenvectors corresponding to the top K largest eigenvalues as the new reduced space.

[0095] Specifically, the dual-channel transformer network-convolutional block attention module DCTN-CBAM architecture consists of a dual-channel transformer network and a convolutional block attention module. The dual-channel transformer network: one channel is used to process time-domain and frequency-domain features, and the other channel is used to extract degradation information from the reduced time-frequency representation TFR. The convolutional block attention module includes a channel attention module and a spatial attention module. The processing process of the convolutional attention module is as Figure 3 shown.

[0096] Specifically, the channel attention module is expressed as:

[0097]

[0098] where W0 and W1 are weight matrices, are the features after average pooling and max pooling, S is the input feature, and S′ is the feature map after the channel attention module.

[0099] Specifically, the spatial attention module is expressed as:

[0100]

[0101] In the formula, M s refers to performing spatial attention module operations on the input features, and f n×n (·) represents a convolution operation with a kernel size of n, represents the stacked feature maps.

[0102] The process of FPT detection includes: determining the FPT based on the bearing HIs, and constructing a monitoring interval for the health state using 3σ. The FPT is defined as follows:

[0103] |HI t+i -m| > 3σ s (i = 0, 1, 2)

[0104] In the formula, HI t+i is the (t + i)-th value of the health index HI, and m and σ s are the mean and standard deviation of the health state monitoring interval, respectively.

[0105] Step 3: Establish a health index degradation model based on the non-linear Wiener process, and combine the failure threshold to achieve high-precision prediction of the remaining life of the bearing.

[0106] (1) For the non-linear characteristics of the rolling bearing health index HI, the non-linear Wiener process is used to characterize the degradation characteristics of the health index HI; let {X(t), t > 0} be the degradation process of the health index HI, and the non-linear Wiener process is defined as follows:

[0107] X(t) = X(0) + at b + σ B B(t)

[0108] In the formula, a is the drift parameter, σ B is the diffusion parameter, b is the unknown parameter in the non-linear function of the drift coefficient, and B(t) represents the standard Brownian motion process reflecting the stochastic dynamics of the degradation process and the inherent time-varying uncertainty used to describe the performance degradation process.

[0109] (2) Define a failure threshold in advance. The life of the rolling bearing ends when its health index HI exceeds the failure threshold, and the probabilistic remaining useful life is predicted. The process is as follows:

[0110] First, define the failure threshold C, which is defined as:

[0111] T = inf{t > 0: X(t) ≥ C|t > 0}

[0112] Wherein, T is the remaining service life, defined as the time when the health index first exceeds the failure threshold C;

[0113] Then calculate the probability distribution of the time when the equipment degradation amount first reaches the failure threshold (i.e., the remaining life), and its expression is:

[0114]

[0115] Wherein, σ B is the diffusion parameter, b is the unknown parameter in the drift coefficient non-linear function, C is the failure threshold, σ a is the standard deviation of the drift parameter a, and μ a is the mean value of the drift parameter a.

[0116] The RUL of the rolling bearing at the k-th observation time is denoted by L k and is expressed as:

[0117] L k = inf{l k > 0: X(l k + t k ) ≥ C|X(t k ) < C};

[0118] Wherein, t k is the time point of the k-th observation, and L k is the remaining service life at the time t k of the k-th observation.

[0119] In order to accurately evaluate the prediction accuracy of the proposed new health index, in the experiment, the new health index and a single health index are respectively used for life prediction, and the mean absolute error is selected as the index of the prediction model accuracy.

[0120] Since the horizontal vibration signal contains richer degradation information, vibration sensors are installed in the horizontal direction to sample the vibration signal. Five rolling bearings of the same model are respectively operated under different working conditions, and the two types of health indexes are respectively used for prediction, and the mean absolute error is calculated to measure the difference between the predicted value and the actual value.

[0121] The 5 types of working conditions are as follows:

[0122] Operating condition type Voltage (V) Load (N) Operating condition 1 100 0 Operating condition 2 100 25 Operating condition 3 100 50 Operating condition 4 100 75 Operating condition 5 100 100

[0123] The above has described in detail an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered as limiting the implementation scope of the present invention. All equivalent changes and improvements made according to the scope of the present invention application should still fall within the patent coverage scope of the present invention.

Claims

1. A method for predicting the bearing life based on a new health index, characterized in that, The prediction method includes the following steps: S1. Collect the vibration signals of the rolling bearing using a vibration sensor, and preprocess the vibration signals; S2. Extract multi-domain features from the preprocessed vibration signals to construct a health index of the bearing; S3. Establish a health index degradation model, set a failure threshold, and then determine the prediction of the remaining life of the bearing according to the probability distribution of the time when the equipment degradation amount first reaches the failure threshold.

2. The method for predicting the bearing life based on the new health index according to claim 1, wherein In step 1, preprocessing the vibration signals is to process the vibration signals by means of continuous wavelet transform, and the expression is: where: \(x(t)\) represents a one-dimensional vibration signal, denotes the complex conjugate operation, \(a\) is the scaling factor, \(b\) is the translation factor, \(\psi\) a,b is the mother wavelet, \(t\) is time, and \(X(a, b)\) is the time-frequency representation TFR used.

3. The method for predicting the bearing life based on the new health index according to claim 1, characterized in that, In step S2, the process of extracting multi-domain features from the preprocessed vibration signals to construct a health index of the bearing includes: S2-1: Use complete ensemble empirical mode decomposition with adaptive noise and kernel principal component analysis to obtain the non-linear health index HI of the bearing; S2-2: Construct multiple health indices HIs in the dual-channel transformer network-convolutional block attention module DCTN-CBAM architecture according to the multi-domain features of the bearing; S2-3: Use the constructed multiple health indices HIs of the bearing for the first prediction time FPT detection.

4. The method for predicting the bearing life based on the new health index according to claim 3, characterized in that, The complete ensemble empirical mode decomposition process with adaptive noise described in step S2-1 includes: (1) Construct an adaptive noise sequence and add it to the original signal to generate multiple groups of random test signals; (2) Perform complete ensemble empirical mode decomposition with adaptive noise CEEMD on each group of random test signals to obtain a group of intrinsic mode function IMF functions; (3) Combine and weight each group of intrinsic mode function IMF functions to generate a group of total intrinsic mode function IMF functions; (4) Perform complete ensemble empirical mode decomposition with adaptive noise CEEMD decomposition on the total intrinsic mode function IMF functions to obtain a group of new intrinsic mode function IMF functions; (5) Check whether the newly generated intrinsic mode function IMF functions meet the convergence condition. If they do not meet the convergence condition, return to step (1) to adjust the adaptive noise; (6) If the convergence condition is met, then repeat steps (2) to (5) for iteration until the number of obtained IMF functions no longer increases.

5. The method for predicting the bearing life based on the new health index according to claim 3, characterized in that The kernel principal component analysis process described in step S2-1: (1) Calculate the kernel matrix K: K ij = k(x i , x j ); where x i and x j are both sample points in the dataset; (2) Centralize the kernel matrix: K′ = K - 1 n K - K1 n +1 n K1 n where 1 n is a matrix of all 1s used to eliminate the influence of the mean; (3) Calculate the eigenvalues and eigenvectors of the kernel matrix: K′v i = λ i v i ; where λ i is the eigenvalue of the kernel matrix, and v i is the corresponding eigenvector; (4) Select the eigenvectors corresponding to the first K largest eigenvalues as the new reduced space.

6. The method for predicting the bearing life based on the new health index according to claim 3, wherein, The dual-channel transformer network-convolutional block attention module DCTN-CBAM architecture described in step S2-2 consists of a dual-channel transformer network and a convolutional block attention module. The dual-channel transformer network: one channel is used to process time-domain and frequency-domain features, and the other channel is used to extract degradation information from the reduced time-frequency representation TFR. The convolutional block attention module includes a channel attention module and a spatial attention module.

7. The method for predicting the bearing life based on the new health index according to claim 6, wherein, The channel attention module is expressed as: Where, W0 and W1 are weight matrices, are the features after average pooling and max pooling, S is the input feature, S′ is the feature map after the channel attention module, and M c refers to performing the channel attention module operation on the input feature.

8. The method for predicting the bearing life based on the new health index according to claim 6, wherein The spatial attention module is expressed as: Wherein, M s refers to performing a spatial attention module operation on the input features, and f n×n (·) represents a convolution operation with a kernel size of n, represents the stacked feature maps.

9. The method for predicting the bearing life based on the new health index according to claim 3, wherein, The process of the first prediction time FPT detection described in step S2-3 includes: using the multiple health indices HIs of the bearing to determine the first prediction time FPT, and using 3σ to construct a monitoring interval for the health state. The first prediction time FPT is defined as follows: |HI t+i -m| > 3σ s (i = 0, 1, 2) where HI t+i is the (t + i)-th value of the health index HI, and m and σ s are the mean and standard deviation of the health status monitoring interval, respectively.

10. The method for predicting the bearing life based on the new health index according to claim 1, characterized in that, In step S3, the process of establishing a health index degradation model, setting a failure threshold, and then determining the prediction of the remaining life of the bearing according to the probability distribution of the time when the equipment degradation amount first reaches the failure threshold includes: (1) For the non-linear characteristics of the rolling bearing health index HI, a non-linear Wiener process is used to characterize the degradation characteristics of the health index HI; let {X(t), t > 0} be the degradation process of the health index HI, and the non-linear Wiener process is defined as follows: X(t) = X(0) + at b + σ B B(t) where a is the drift parameter, σ B is the diffusion parameter, b is the unknown parameter in the nonlinear function of the drift coefficient, and B(t) represents the standard Brownian motion process reflecting the stochastic dynamics of the degradation process and the inherent time-varying uncertainty used to describe the performance degradation process; (2) A failure threshold is predefined in advance. The life of the rolling bearing ends when its health index HI exceeds the failure threshold, and the probability prediction of the remaining service life is carried out as follows: First, define the failure threshold C, which is defined as: T = inf{t > 0: X(t) ≥ C|t > 0} In the formula, T is the remaining service life, which is defined as the time when the health index first exceeds the failure threshold C; Then calculate the probability distribution of the time when the equipment degradation amount first reaches the failure threshold, and its expression is: where σ B is the diffusion parameter, b is the unknown parameter in the nonlinear function of the drift coefficient, C is the failure threshold, σ a is the standard deviation of the drift parameter a, and μ a is the mean of the drift parameter a; The remaining useful life RUL of the rolling bearing at the k-th observation time is denoted by L k and expressed as: L k = inf{l k > 0: X(l k + t k ) ≥ C | X(t k ) < C} where t k is the time point of the k-th observation, and L k is the remaining useful life at the time t k of the k-th observation.

Images