A method for predicting rolling bearing reliability

By combining data-driven and model-driven methods, using Spearman correlation coefficient, random forest algorithm and BiLSTM neural network, a weighted health index curve and stress-strength model are constructed, which solves the volatility of rolling bearing reliability prediction in the existing technology and Human factors influence, high accuracy and high efficiency reliability prediction is achieved.

CN118194461BActive Publication Date: 2024-11-15TIANJIN POLYTECHNIC UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410339173.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-25
Publication Date
2024-11-15
Estimated Expiration
2044-03-25

AI Technical Summary

Technical Problem

In the reliability prediction of rolling bearings, the prior art fails to effectively combine the vibration data characteristics and reliability, resulting in large fluctuations in the prediction results and human factors, which is difficult to reflect the actual situation.

Method used

Using a method combining data-driven and model-driven, by collecting full-life vibration data, extracting time-domain and frequency-domain features, using Spearman correlation coefficient and random forest algorithm to screen features, construct weighted health index curves, and establishing stress- The intensity interference model, combined with Bayesian optimization algorithm and BiLSTM neural network, predicts the reliability of rolling bearings.

Benefits of technology

It improves the accuracy and accuracy of rolling bearing reliability prediction, reduces the influence of human factors, and provides a high accuracy and high efficiency prediction model, which can better reflect the health status and reliability of the bearing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118194461B_ABST
    Figure CN118194461B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for predicting the reliability of a rolling bearing. First, multi-domain feature extraction is performed on the horizontal vibration data of the rolling bearing to construct a multi-domain high-dimensional feature set; the Spearman correlation coefficient is then used to screen out time-independent features in the multi-domain high-dimensional feature set, and the random forest algorithm is used to eliminate redundant features to form a low-dimensional feature set; then a weighted HI curve is constructed using the Box‑Cox transformation and the weighted coefficient method; then the gradual reliability curve of the rolling bearing is obtained according to the random perturbation method and the fourth-order moment method; finally, the weighted HI curve is used as the input of the BOA‑BiLSTM prediction model, and the gradual reliability curve is used as the label of the BOA‑BiLSTM prediction model to predict the reliability of the target rolling bearing. The present invention adopts the method of adding labels to the prediction model for the first time to perform reliability prediction, and predicts the reliability of the target rolling bearing based on health indicators, gradual reliability and the BOA‑BiLSTM prediction model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of rolling bearing reliability assessment, and in particular is a rolling bearing reliability prediction method. Background Art

[0002] As a key component in the mechanical field, rolling bearings work by utilizing the rolling of rolling elements between the inner and outer rings, thereby reducing friction and energy loss. This design can withstand large loads while having low friction and wear. Therefore, rolling bearings are often used in aerospace, mechanical equipment, and vehicles, and their health status directly affects safe production. Therefore, health assessment and reliability prediction of rolling bearings are of great research significance.

[0003] At present, two research methods, linear degradation and segmented degradation, are usually used for health assessment of rolling bearings. These two methods greatly improve the accuracy of bearing failure time prediction, but do not take into account the impact of actual factors such as the bearing working environment and load on the bearing life. Therefore, people process the bearing vibration signal and describe the operation of the system by establishing health indicators. Due to the complex degradation state of the bearing, the use of a single feature cannot fully reflect the degradation information of the bearing. Therefore, people began to extract multi-domain high-dimensional feature sets of the full-life vibration signal and use PCA, KPCA, LLE and other methods to reduce the dimension, reduce the calculation process, and improve the accuracy of the prediction.

[0004] When selecting multiple feature indicators for fusion to construct the HI (Health Indicator) curve, it is necessary to use prior knowledge and manual experience to extract multi-domain features. The rationality and accuracy of the constructed health indicators have a great influence on the accuracy of the remaining service life prediction results of the bearing. Therefore, the prediction of the health status of the bearing through health indicators is greatly affected by human factors.

[0005] Reliability is based on statistics and probability theory, and is studied by establishing a reliability model based on historical degradation data. This research method pays more attention to the physical characteristics of the system, environmental factors and other practical factors, and is relatively less affected by human factors. In order to reduce downtime and maintenance costs, extend equipment life, and meet the mission reliability requirements of the equipment, it is necessary to accurately evaluate the reliability of the bearing.

[0006] According to different methods, the prediction of bearing reliability can be divided into two categories, namely model-driven and data-driven methods. The model-driven method is one of the common methods for bearing reliability assessment and prediction. This method describes the performance characteristics of the bearing by establishing a mathematical model, and performs reliability assessment and prediction. However, this method usually causes large fluctuations in the reliability curve, which is different from the actual situation. The data-driven method is based on statistical knowledge and machine learning technology, and usually includes steps such as data collection, feature extraction, degradation assessment, and remaining life prediction. This method does not require knowledge of the physical degradation mechanism of the bearing. It explores the degradation law of the rolling bearing through data analysis and then predicts its remaining service life.

[0007] In previous studies, vibration data characteristics and reliability have rarely been combined with each other. Reliability prediction using a combination of data-driven and model-driven methods is difficult to reflect the relationship between vibration characteristics and reliability. Summary of the invention

[0008] In view of the deficiencies in the prior art, the technical problem to be solved by the present invention is to provide a rolling bearing reliability prediction method.

[0009] The technical solution of the present invention to solve the technical problem is to provide a rolling bearing reliability prediction method, characterized in that the method comprises the following steps:

[0010] Step 1, collecting the full life vibration data of the rolling bearing; the full life vibration data of the rolling bearing includes the horizontal vibration data of the rolling bearing and the vertical vibration data of the rolling bearing;

[0011] Step 2, extracting the time domain features and frequency domain features of the rolling bearing horizontal vibration data obtained in step 1 to form a multi-domain high-dimensional feature set;

[0012] Step 3: First, use the Spearman correlation coefficient to filter out irrelevant features in the multi-domain high-dimensional feature set, and then use the random forest algorithm to eliminate redundant features in the multi-domain high-dimensional feature set to form a low-dimensional feature set;

[0013] Step 4: Use Box-Cox transformation to fuse the low-dimensional feature set obtained in step 3 to obtain a HI curve related to the deterioration state of the rolling bearing; then reconstruct the HI curve according to the weighted coefficient method to obtain a weighted HI curve;

[0014] Step 5, select the basic random variable X that affects the reliability of the rolling bearing, then establish the stress-strength interference model, then establish the gradual reliability state function of the rolling bearing according to the Gamma degradation process based on the stress-strength interference model, and then obtain the gradual reliability curve of the rolling bearing according to the random perturbation method and the fourth-order moment method;

[0015] Step 6: Use the Bayesian optimization algorithm to optimize the hyperparameters of the BiLSTM neural network to obtain a BOA-BiLSTM prediction model with optimal hyperparameters; then use the weighted HI curve obtained in step 4 as the input of the BOA-BiLSTM prediction model and the gradient reliability curve obtained in step 5 as the label of the BOA-BiLSTM prediction model to predict the reliability of the target rolling bearing.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] (1) The present invention adopts the method of adding labels to the prediction model for the first time to perform reliability prediction, and predicts the reliability of the target rolling bearing based on health indicators, gradual reliability and BOA-BiLSTM prediction model.

[0018] (2) The present invention, for the first time, takes the gradual reliability taking into account the operating parameters and the random parameters of the bearing as the label of the prediction model, and uses a hybrid driving method combining data-driven and model-driven to perform reliability prediction, so as to improve the reliability prediction accuracy.

[0019] (3) The present invention combines the bearing vibration characteristics and health status to propose a data-driven reliability prediction method based on gradual degradation mechanism and health indicators.

[0020] (4) The present invention proposes for the first time a highly accurate and efficient BOA-BiLSTM prediction model for bearing reliability evaluation and prediction.

[0021] (5) This invention proposes for the first time a method for constructing a bearing HI curve based on Box-Cox transformation and weighting coefficients to improve the accuracy of characterizing bearing degradation information.

[0022] (6) The present invention uses the random forest algorithm for the first time to screen the degradation characteristics of bearings, solving the problem that principal component analysis is difficult to achieve nonlinear data dimensionality reduction and kernel principal component analysis is difficult to calculate. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is the overall flow chart of the present invention;

[0024] Figure 2 are 16 time domain features of Example 1 of the present invention;

[0025] Figure 3 are 12 frequency domain features of Embodiment 1 of the present invention;

[0026] Figure 4 This is an importance score diagram of the time domain features after the Spearman correlation coefficient is removed in Example 1 of the present invention;

[0027] Figure 5This is an importance score graph of frequency domain features after the Spearman correlation coefficient is removed in Example 1 of the present invention;

[0028] Figure 6 This is a waveform diagram of the time domain features after redundant features are eliminated by the random forest algorithm according to Example 1 of the present invention;

[0029] Figure 7 This is a waveform diagram of frequency domain features after redundant features are eliminated by the random forest algorithm according to Example 1 of the present invention;

[0030] Figure 8 is a weighted HI curve diagram of Example 1 of the present invention;

[0031] Fig. 9 is a gradual reliability curve diagram of Example 1 of the present invention;

[0032] Fig.10 A logistic reliability curve obtained by using an existing logistic regression model according to Example 1 of the present invention;

[0033] Fig.11 A sensitivity diagram of the basic random variable X to the mean value in Example 1 of the present invention;

[0034] Fig.12 A sensitivity diagram of the basic random variable X to the variance in Example 1 of the present invention;

[0035] Fig.13 This is a prediction result diagram of the BOA-BiLSTM prediction model of Example 1 of the present invention;

[0036] Fig.14 This is a prediction result diagram of the SSA-BiLSTM prediction model of Example 1 of the present invention;

[0037] Fig.15 This is a diagram of the prediction results of the BiLSTM prediction model of Example 1 of the present invention. DETAILED DESCRIPTION

[0038] The specific embodiments of the present invention are given below. The specific embodiments are only used to further illustrate the present invention in detail and do not limit the protection scope of the claims of the present invention.

[0039] The present invention provides a rolling bearing reliability prediction method (hereinafter referred to as the method), characterized in that the method comprises the following steps:

[0040] Step 1, collecting the full life vibration data of the rolling bearing; the full life vibration data of the rolling bearing includes the horizontal vibration data of the rolling bearing (i.e., the vibration data in the horizontal direction) and the vertical vibration data of the rolling bearing (i.e., the vibration data in the vertical direction);

[0041] Preferably, in step 1, the present invention uses the existing XJTU-SY rolling bearing accelerated life test data set for research.

[0042] Step 2: extracting the time domain features and frequency domain features of the rolling bearing horizontal vibration data obtained in step 1 by a feature extraction method to form a multi-domain high-dimensional feature set;

[0043] Preferably, in step 2, the time domain characteristics include 16 time domain characteristics with expressions, namely, mean, minimum value, root mean square, peak-to-peak value, root square amplitude, waveform index, absolute average value, peak index, kurtosis, margin index, variance, skewness index, maximum value, kurtosis index, skewness and pulse index.

[0044] Preferably, in step 2, the frequency domain features include 12 frequency domain features with expressions, namely, spectral mean, spectral root mean square value, frequency domain amplitude skewness index, frequency domain amplitude kurtosis index, centroid frequency, frequency variance, root mean square frequency, amplitude ratio index, amplitude variance ratio index, frequency density index, frequency domain frequency skewness and frequency domain frequency kurtosis.

[0045] Step 3: First, use the Spearman correlation coefficient to filter out irrelevant features in the multi-domain high-dimensional feature set, and then use the random forest algorithm to eliminate redundant features in the multi-domain high-dimensional feature set to form a low-dimensional feature set, thereby achieving dimensionality reduction and improving prediction accuracy;

[0046] Preferably, the specific steps of step 3 are as follows:

[0047] S3.1. Calculate the correlation between time and each feature in the multi-domain high-dimensional feature set through the Spearman correlation coefficient, and then filter out the features whose correlation with time is not higher than the set correlation coefficient threshold (i.e., irrelevant features) to obtain the time-related feature set;

[0048] Preferably, in step S3.1, the correlation coefficient threshold is set to 0.7; 0≤r<0.2 means no correlation, 0.2≤r≤0.4 means weak correlation, 0.4<r≤0.6 means moderate correlation, 0.6<r≤0.8 means strong correlation, and 0.8<r≤1.0 means extremely strong correlation.

[0049] S3.2. Calculate the importance score of each feature in the time-related feature set obtained in step S3.1 through the random forest algorithm, and then sort the importance scores from large to small, retaining the features whose importance scores are not less than the set importance score threshold, thereby eliminating redundant features and forming a low-dimensional feature set.

[0050] Preferably, in step S3.2, the importance score threshold is set to 0.5.

[0051] Preferably, in step S3.2, calculating the importance score of the feature by the random forest algorithm is achieved by Gini Importance, and the specific steps are as follows:

[0052] S3.2.1. Calculate Gini impurity G(i): Gini impurity G(i) is an indicator for measuring data purity in decision tree algorithms. Random forest algorithm is a type of decision tree algorithm. In the process of building a decision tree, Gini impurity G(i) is used to measure the degree of mixing of samples of different categories in a node i. The lower the Gini impurity G(i), the more samples in node i tend to belong to the same category j. The higher the sample purity, the more credible the decision made at the node.

[0053] The calculation formula of Gini impurity G(i) is:

[0054]

[0055] In formula (1), N is the number of categories; p j is the probability that the sample belongs to the jth category.

[0056] S3.2.2. Calculate Gini importance GI(i): Each node i in the random forest algorithm will consider splitting a part of the features. Gini importance GI(i) is used to measure the degree of reduction of Gini impurity G(i) after splitting through node i, that is, the reduction of Gini impurity G(i); the greater the reduction, the more important this split is to improve the feature importance score;

[0057] The calculation formula of the Gini importance GI(i) of each node is:

[0058] GI(i)=w i ΔG(i) (2)

[0059] In formula (2), w i is the sample weight in node i; ΔG(i) is the change in Gini impurity before and after the split of node i.

[0060] S3.2.3. Calculate the importance score FI of each feature: Take the weighted average of the Gini importance GI(i) of all nodes i included in each feature to obtain the importance score FI of each feature, as shown in formula (3):

[0061]

[0062] In formula (3), N represents the number of nodes for each feature; w i is the sample weight of node i, n i is the number of samples.

[0063] Step 4: Use Box-Cox transformation to fuse the low-dimensional feature set obtained in step 3 to obtain a HI curve related to the deterioration state of the rolling bearing; then reconstruct the HI curve according to the weighted coefficient method to obtain a weighted HI curve related to the deterioration state of the rolling bearing that is more in line with reality;

[0064] Preferably, the specific steps of step 4 are as follows:

[0065] S4.1. In order to perform feature fusion, it is assumed that the importance of each feature to health degradation is the same, and then the low-dimensional feature set obtained in step 3 is fused using Box-Cox transformation to obtain the HI curve related to the degradation state of the rolling bearing (hereinafter referred to as HI curve);

[0066] Preferably, in step S4.1, the Box-Cox transformation is a statistical method for data transformation, which can adjust the distribution form of data to make it closer to normal distribution; the form of the Box-Cox transformation is:

[0067]

[0068] In formula (4), y is the original data; λ is the transformation parameter.

[0069] 4.2. Since the degradation information represented by each feature in the low-dimensional feature set obtained in step 3 is different, the Spearman correlation coefficient between each feature and time obtained in step S3.1 is used as the weight of the corresponding feature, and then all weights are standardized (in this embodiment, the standardization adopts MinMax scaling normalization) to between 0 and 1, and then the HI curve obtained in step S4.1 is reconstructed according to the standardized weights according to the weighted coefficient method to obtain a weighted HI curve that is more in line with reality and related to the degradation state of the rolling bearing.

[0070] Step 5, select the basic random variable X that affects the reliability of the rolling bearing, then establish the stress-strength interference model, then establish the gradual reliability state function of the rolling bearing according to the Gamma degradation process based on the stress-strength interference model, and then obtain the gradual reliability curve of the rolling bearing according to the random perturbation method and the fourth-order moment method;

[0071] Preferably, the specific steps of step 5 are as follows:

[0072] S5.1. Select the basic random variables X that affect the reliability of rolling bearings;

[0073] S5.2. Establish a gradual reliability state function for rolling bearings;

[0074] Since bearing components are affected by factors such as temperature, humidity, and load, their strength will gradually deteriorate over time. As the strength of bearing components gradually decreases, the strength distribution will inevitably intersect with the stress distribution. There is a dynamic relationship between stress and strength, which leads to the stress-strength interference model, which treats stress and strength as random variables. Based on the stress-strength interference model, the gradual reliability state function of the rolling bearing is established according to the Gamma degradation process:

[0075] g(t,X)=r(t)-σ(t,X) (5)

[0076] In formula (5), r(t) is the random degradation process of component strength at time t; σ(t,X) is the process of random load action, and X is the basic random variable affecting the reliability of rolling bearings; the random degradation process of component strength at time t r(t) is:

[0077] r(t)=r 0 -Z(t) (6)

[0078] In formula (6), r 0 is the initial strength; Z(t) is the strength degradation process of the part at time t in the Gamma degradation process; z is the strength degradation amount of the part from time 0 to time t, and its probability density function is:

[0079]

[0080] In formula (7), when z∈[0,∞], I 0,∞ (z) is 1, when Time I 0,∞ (z) is 0; v is the shape parameter; u is the size parameter; is the Gamma function;

[0081] The mean E[Z(t)] and variance Var[Z(t)] of the probability density function are:

[0082]

[0083]

[0084] In formula (8)-formula (9), b, c, u are the parameters of the Gamma degradation process, and their values ​​are obtained by the maximum likelihood estimation and moment estimation of statistical data;

[0085] S5.3. According to the random perturbation method and the fourth-order moment method, the gradual reliability curve of the rolling bearing is obtained:

[0086] The basic gradual reliability model of rolling bearings is obtained from equations (5) and (6):

[0087]

[0088] When the basic random variable distribution of the rolling bearing obeys the normal distribution and the movement of the bearing is relatively stable, the reliability index β and reliability R(t) obtained according to formula (10) and the first-order second moment are:

[0089]

[0090] R(t)=Φ[β(t)]=P{r(t)>σ(t,X),t∈[0,T s ]} (12)

[0091] In formula (11) and formula (12), μ g(t) , σ g(t) is the first two moments of the state function at time t; E[X] and Var(X) are the mean and variance of the basic random variables; Φ[X] is the standard normal cumulative distribution function; T s is the limit state moment;

[0092] Under high speed and heavy load conditions, as the bearing is used for a long time, its strength gradually deteriorates and its performance continues to decrease, causing nonlinear responses such as friction changes and contact deformation. In this case, the second-order moment method is not applicable to solve the reliability. In order to calculate the gradual reliability of rolling bearings, random perturbation theory and the fourth-order moment method are used to solve the gradual reliability of rolling bearings based on equations (11) and (12);

[0093] According to the random perturbation theory, by performing Taylor series expansion on the limit state function, the probability characteristics of the limit state function are obtained as follows:

[0094]

[0095]

[0096]

[0097]

[0098] In formulas (13) to (16), is the (·)Kronecker power, symbol represents the Kronecker product, defined as μ g , θ g and η g is the first four moments of the limit state function; C 3 (X) and C 4 (X) is the third and fourth order moments of the basic random variable;

[0099] When the distribution probability cannot be determined, but there is sufficient data to determine the first four moments of the design parameters, the probability distribution function of the standardized random variable that obeys an arbitrary distribution is approximately expanded into a standard normal distribution function according to the Edgeworth series method as shown in formula (17):

[0100]

[0101] In formula (17), Φ(.) is the standard normal cumulative distribution function; is the standard normal probability density function; H j (y) is the j-order Hermite polynomial, and its recursive relation is:

[0102]

[0103] According to equations (17) and (18), the improved gradual reliability model R(β(t)) of rolling bearing is obtained as follows:

[0104] R(β(t))=P[g(t,X)>0]=1-F(-β(t)) (19)

[0105] When the Edgeworth series method is used to take the first four items of the standard normal distribution function F(-β(t)) to calculate the gradual reliability, R>1 may occur. When this happens, the empirical correction formula is used to process it, so that the calculation result is more accurate, and the gradual reliability curve of the rolling bearing is finally obtained.

[0106] The experience correction formula is:

[0107]

[0108] Preferably, in step 5, the sensitivity of the mean and variance of the basic random variable X reflects the influence of the basic random variable X on the gradual reliability. Therefore, the influence of the basic random variable X selected in step S5.1 on the gradual reliability is determined by the results obtained by equations (21) and (22);

[0109]

[0110]

[0111] Step 6. Use the Bayesian optimization algorithm to optimize the hyperparameters of the BiLSTM neural network to obtain the BOA-BiLSTM (Bayesian optimized bidirectional long short-term memory neural network) prediction model with the optimal hyperparameters; then use the weighted HI curve obtained in step 4 as the input of the BOA-BiLSTM prediction model and the gradual reliability curve obtained in step 5 as the label of the BOA-BiLSTM prediction model to predict the reliability of the target rolling bearing.

[0112] Preferably, in step 6, the hyperparameters of the BiLSTM neural network include an initial learning rate, a regularization parameter, a maximum number of iterations, and the number of hidden layer nodes.

[0113] Preferably, in step 6, the specific steps of establishing a BOA-BiLSTM prediction model with optimal hyperparameters are as follows:

[0114] S6.1. Establishing Gaussian process model: For any x∈X, X is the input space, the Gaussian process is expressed as:

[0115] f(x)~GP(m(x),k(x,x′)) (23)

[0116] In formula (23), f(x) is the objective function; m(x) is the mean function of the Gaussian process model; k(x, x′) is the covariance function;

[0117] S6.2. Select sampling strategy: The sampling strategy is a method for selecting the next point to be evaluated. The present invention adopts the high confidence upper bound (UCB) as the sampling strategy, which is expressed as:

[0118]

[0119] In formula (24), β is a parameter that weighs exploration and exploitation;

[0120] S6.3, evaluate the objective function: pass the selected points as input to the objective function, and calculate the function value for the selected points;

[0121] S6.4, update the Gaussian process model: merge the function value calculated in step S6.3 into the Gaussian process model, and recalculate the mean function m(x) and covariance function k(x, x′) of the Gaussian process model;

[0122] S6.5. Output optimal hyperparameters: Stop when the set number of iterations, time limit or other termination conditions are reached, and return the point currently considered to be the optimal solution.

[0123] Example 1

[0124] In step 1, this embodiment uses the XJTU-SY rolling bearing accelerated life test data set for research. The data set Bearing1_1 contains 4030464 sampling points, the sampling frequency is 25600Hz, and a total of 123 sampling acquisitions are performed, each acquisition lasts 1.28 seconds, and the acquisition time interval is 60 seconds. The test bearing is a LDKUER204 rolling bearing. Since a force of 12KN is applied to the bearing in the horizontal direction, the horizontal vibration data of the rolling bearing contains more degradation information, so the horizontal vibration data of the rolling bearing is selected for analysis.

[0125] In step 2, the 16 time domain features extracted are as follows Figure 2 As shown, the 12 frequency domain features are as follows Figure 3 shown.

[0126] In step 3, the correlation coefficient threshold is set to 0.7. After calculation and combining with the set correlation coefficient threshold, seven time domain features, namely, root mean square (RMS), root amplitude (RA), absolute mean (AM), kurtosis (K), variance (Var), maximum value (MAX) and peak-to-peak value (PP), are retained in the time domain features, and the spectrum mean (P 1 ), spectrum RMS value (P 2 ), frequency variance (P 6 ), frequency density index (P 10 ) and frequency kurtosis (P 12 )5 frequency domain features.

[0127] The importance score threshold is set to 0.5. The importance scores of the 7 time domain features and 5 frequency domain features retained by the Spearman correlation coefficient are as follows: Figure 4 and Figure 5 As shown. Figure 4-Figure 5 It can be seen that after the random forest algorithm, two time domain features of peak-to-peak value and kurtosis and three frequency domain features of spectrum mean, spectrum root mean square value and frequency domain frequency kurtosis are retained.

[0128] like Figure 6-Figure 7 As shown in the figure, the amplitude variation trends of the two retained time domain features and the three frequency domain features are consistent, which can reasonably characterize the degradation process of the rolling bearing.

[0129] In step 4, the weighted HI curve is as follows Figure 8 As shown. Figure 8 It can be seen that the rolling bearing degrades slowly from the beginning of the test to 624, and then the health state changes suddenly from 624, and then the health degradation rate is faster. When it reaches 946, it degrades rapidly until failure. The weighted HI curve is consistent with the amplitude change trend of the original horizontal vibration signal, which can reasonably characterize the degradation process of the rolling bearing.

[0130] In step 5, the basic random variables X include the inner ring diameter, outer ring diameter, bearing median diameter, rotation speed and rolling element diameter, and the relevant statistical parameters are shown in Table 1;

[0131] Table 1

[0132] random variable Variable units Distribution Probability First Moment Second moment Third-order moment Fourth-order moment Inner ring diameter mm normal distribution 29.30 0.0488 0 <![CDATA[1.706e -05 ]]> Outer ring diameter mm normal distribution 39.80 0.0663 0 <![CDATA[5.808e -05 ]]> Bearing diameter mm normal distribution 34.55 0.0576 0 <![CDATA[3.298e -05 ]]> Speed r / min normal distribution 2100 3.5 0 450.19 Rolling element diameter mm normal distribution 7.92 0.0132 0 <![CDATA[9.108e -08 ]]>

[0133] In step 5, the gradient reliability curve is obtained as Fig. 9 As shown. Fig. 9 It can be seen that the reliability of the bearing gradually decreases with the increase of service time. In the early stage of use, the reliability decreases slowly, and as the service time increases, the reliability decreases faster and faster. Fig.10 The trend is consistent with Fig. 9 The reliability decreases monotonically, which is more in line with reality.

[0134] Depend on Figure 11-Figure 12 It can be seen that among the basic random variables X, the sensitivity of the mean and variance of the outer ring diameter are the largest, indicating that the change of the outer ring diameter has the greatest impact on the gradual reliability of the bearing. The mean of the outer ring diameter has a negative impact on the gradual reliability, indicating that the larger the outer ring diameter of the rolling bearing, the smaller the bearing reliability and the less reliable the bearing. The variance of the outer ring diameter has a positive impact on the gradual reliability.

[0135] In step 6, the 984 sample data are divided into training set and test set, and the BOA-BiLSTM prediction model is used for prediction. The prediction results are as follows: Fig.13 shown.

[0136] In order to verify the accuracy of BOA-BiLSTM, this example compares the prediction results of BOA-BiLSTM with those of BilSTM (bidirectional long short-term memory neural network) and SSA-LSTM (bidirectional long short-term memory neural network optimized by sparrow algorithm), and uses Bearing1_1 for verification. Figure 13-Figure 15 It can be seen that the prediction results of BOA-BiLSTM are closer to the true value, and the BOA-BiLSTM prediction model has better prediction performance.

[0137] Table 2

[0138]

[0139]

[0140] Table 2 is a comparison of the errors of various prediction models. It can be seen from Table 2 that the RMSE, MAE, and MAPE of the BOA-BiLSTM prediction model of the present invention on the Bearing1_1 test set are small, and the calculation speed is greatly improved compared with the SSA-BiLSTM, which shows that the prediction accuracy of BOA-BiLSTM is high and the performance is good.

[0141] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A rolling bearing reliability prediction method, characterized in that: The method comprises the following steps: Step 1, collecting the full life vibration data of the rolling bearing; the full life vibration data of the rolling bearing includes the horizontal vibration data of the rolling bearing and the vertical vibration data of the rolling bearing; Step 2, extracting the time domain features and frequency domain features of the rolling bearing horizontal vibration data obtained in step 1 to form a multi-domain high-dimensional feature set; Step 3: First, use the Spearman correlation coefficient to filter out irrelevant features in the multi-domain high-dimensional feature set, and then use the random forest algorithm to eliminate redundant features in the multi-domain high-dimensional feature set to form a low-dimensional feature set; The specific steps of step 3 are as follows: S3.

1. Calculate the correlation between time and each feature in the multi-domain high-dimensional feature set through the Spearman correlation coefficient, and then filter out the features whose correlation with time is not higher than the set correlation coefficient threshold to obtain the time-related feature set; S3.2, calculating the importance score of each feature in the time-related feature set obtained in step S3.1 by using the random forest algorithm, retaining the features whose importance score is not less than the set importance score threshold, thereby eliminating redundant features and forming a low-dimensional feature set; The specific steps for calculating the importance score of features based on the random forest algorithm are as follows: S3.2.

1. Calculate Gini impurity G(i): Gini impurity G(i) is used to measure the degree of mixing of samples of different categories in a node i. The lower the Gini impurity G(i), the more samples in node i tend to belong to the same category j, and the higher the sample purity. The calculation formula of Gini impurity G(i) is: In formula (1), N is the number of categories; p j is the probability that the sample belongs to the jth category; S3.2.

2. Calculate Gini Importance GI(i): Gini Importance GI(i) is used to measure the reduction of Gini Impurity G(i) after splitting through node i. The larger the reduction, the more important this split is for improving the feature importance score. The calculation formula of the Gini importance GI(i) of each node is: GI(i)=in i ΔG(i) (2) In formula (2), w i is the sample weight in node i; ΔG(i) is the change in Gini impurity before and after the split of node i; S3.2.

3. Calculate the importance score FI of each feature: Take the weighted average of the Gini importance GI(i) of all nodes i included in each feature to obtain the importance score FI of each feature, as shown in formula (3): In formula (3), N represents the number of nodes for each feature; w i is the sample weight of node i; Step 4: Use Box-Cox transformation to fuse the low-dimensional feature set obtained in step 3 to obtain a HI curve related to the deterioration state of the rolling bearing; then reconstruct the HI curve according to the weighted coefficient method to obtain a weighted HI curve; The specific steps of step 4 are as follows: S4.1, assuming that the importance of each feature to health degradation is the same, then use the Box-Cox transformation to fuse the low-dimensional feature set obtained in step 3 to obtain the HI curve related to the degradation state of the rolling bearing; 4.

2. The Spearman correlation coefficient between each feature and time obtained in step S3.1 is used as the weight of the corresponding feature, and all weights are standardized to between 0 and 1. Then, the HI curve obtained in step S4.1 is reconstructed according to the standardized weights according to the weighted coefficient method to obtain a weighted HI curve; Step 5, select the basic random variable X that affects the reliability of the rolling bearing, then establish the stress-strength interference model, then establish the gradual reliability state function of the rolling bearing according to the Gamma degradation process based on the stress-strength interference model, and then obtain the gradual reliability curve of the rolling bearing according to the random perturbation method and the fourth-order moment method; Step 6: Use the Bayesian optimization algorithm to optimize the hyperparameters of the BiLSTM neural network to obtain a BOA-BiLSTM prediction model with optimal hyperparameters; then use the weighted HI curve obtained in step 4 as the input of the BOA-BiLSTM prediction model and the gradient reliability curve obtained in step 5 as the label of the BOA-BiLSTM prediction model to predict the reliability of the target rolling bearing.

2. The rolling bearing reliability prediction method according to claim 1, characterized in that: In step 2, the time domain features include mean, minimum, root mean square, root square amplitude, waveform index, absolute average, peak index, margin index, variance, skewness index, maximum, kurtosis index and pulse index; The frequency domain features include spectrum mean, spectrum root mean square value, frequency domain amplitude skewness index, frequency domain amplitude kurtosis index, center of gravity frequency, frequency variance, root mean square frequency, amplitude ratio index, amplitude variance ratio index, frequency density index, frequency domain frequency skewness and frequency domain frequency kurtosis.

3. The rolling bearing reliability prediction method according to claim 1, characterized in that: In step S4.1, the Box-Cox transformation is in the form of: In formula (4), y is the original data and λ is the transformation parameter.